I design, architect and build software products.
2 product slots open for Q4Systems Architect, Virginia, USA. I design, architect and build software products.
naram@hey.com · www.naramalhasani.com · linkedin.com/in/NaramAlhasani
10 systems, 7 with live sites
Construction project platform. Running project teams on big projects right now: real-time chat, RFIs, submittals and bids in one place, always showing who owes the next move.
In use on live projects
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AClients
UI · Next.js 15 · React 19
The app people use on site and in the office, built mobile-first for phones and iPads. Types are generated from the Java DTOs, so the front end and back end can’t drift.
Talks to
UI · one-time token pages
Pages that outside architects, subs and reviewers open from an emailed link to answer an RFI or a submittal, or to submit a bid. No account needed.
Talks to
EXTERNAL · MCP client
Claude Desktop or any MCP client, signed in as a real user through OAuth. It only gets what that person could do in the app.
Talks to
BGateway
SERVICE · Spring Boot · WebFlux
The only door in. It checks the JWT, resolves the user, their company and the client IP once, and forwards each request with trusted identity headers.
Talks to
REALTIME · WebSockets
Held by the gateway: chat, notifications, presence and which conversation each tab is reading. Drawn in full below.
Talks to
AI · JSON-RPC
The MCP transport, also in the gateway. Claude’s tool calls arrive here and run as the signed-in user.
Talks to
AI · Node · Fastify
Runs the assistant and the agents’ turns: streams the model, loads the tool list, applies each company’s model policy, evaluates agent hooks once per turn, and transcribes voice notes.
Talks to
CPlatform
SERVICE · dashboard · directory
The centre of the model. Everything else belongs to a project and its team, and the dashboard is built around what the job needs today.
Talks to
SERVICE · ball in court
Requests for information, with drafts. Status rules move the ball: open sits with the assignee’s company, answered or returned goes back to the asker, closed clears it.
Talks to
SERVICE · revisions · ball in court
Submittals go round in revisions, with grant-based access, private drafts and a ball held by named people, until one is approved.
SERVICE · packages · leveling
Bid packages per trade, bids with a schedule of values and alternates, leveling, and the award.
Talks to
SERVICE · split · diff · presence
Splits a plan-set PDF into one sheet per page, renders a pixel diff between revisions, holds markups and pins, and shows who else is on the sheet.
Talks to
SERVICE · real time
Project, group and direct conversations with mentions and attachments. Drawn in full below.
Talks to
SERVICE · contracts · SOV · pay apps
Contracts with schedules of values, change events and change orders, and pay applications on the AIA G702/G703 forms. Contract sums are derived, never stored.
Talks to
AI · commands · skills · hooks
Each company’s AI agents: a built-in set of construction commands, its own skills, hooks evaluated on every turn with an audit trail, and delegation to named agents only.
Talks to
SERVICE · in-app · email
Tells the right people when something moves, like a ball changing hands, in the app and by email.
Talks to
AI · 199 MCP tools
Every action an AI may take, declared on the service method itself with a schema derived from its DTO, one error path, and the caller’s permissions checked on every call.
Talks to
SERVICE · library · PDF
One library for files and generated documents at company and project level, each with a share list, rendered to PDF on the server.
Talks to
SERVICE · OAuth 2.1
Lets MCP clients such as Claude Desktop sign in as a user. Redirect addresses are checked strictly so the consent step can’t be abused.
Talks to
SERVICE · leads · estimating
Leads with statuses each company sets, client accounts, contacts, follow-ups, the estimating schedule, and won/lost analysis.
Talks to
JOB · triggers · cron
Workflow definitions wired step to step, checked so each output fits the next input, and run on triggers or a schedule.
Talks to
AI · from past documents
Reads a company’s past project documents and proposes budget and qualification catalogue entries for a person to accept.
DData
STORE · Liquibase migrations
The system of record for every module, with versioned schema migrations.
Talks to
STORE · cache · pub/sub
Caching, plus the pub/sub channels that carry chat, notification, presence and viewing events between instances.
Talks to
STORE · S3-compatible
Files: drawing sheets and thumbnails, diff overlays, photos and documents, served through short-lived signed links.
Talks to
EOutside
EXTERNAL · one-way import
Projects are imported from Procore over OAuth. Nothing is written back.
Talks to
EXTERNAL · webhooks
The shared inbox. Its webhooks bring email activity into sales.
Talks to
EXTERNAL · models
The models behind the assistant and the agents, chosen per company within the platform’s policy.
Talks to
EXTERNAL · OAuth · files
Autodesk Platform Services: linked projects, folders and files.
Talks to
Connections
Inside: real-time chat and presence
Every server can hold someone’s open tab, so nothing is sent tab to tab. Events go through Redis once, each server pushes them to the tabs it holds, and the app keeps track of what each tab is actually looking at.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
ATabs
UI · one socket each
Every open tab is its own WebSocket session. It reports which conversation it is showing and whether the person is active or idle.
Talks to
UI · avatars
Shows who else is on the same sheet right now.
Talks to
BGateway
REALTIME · per-session state
Tracks viewing per session, not per user: when a tab closes it re-resolves from the person’s remaining tabs, so a closed tab never leaves a stale “still reading”.
Talks to
REALTIME · channel subscribers
Subscribes to the chat, notification and system channels and pushes each event to the sessions this server holds.
Talks to
REALTIME · JWT handshake
Relays “I’m on this sheet” to the platform. The user comes from the authenticated handshake, never the message, and every keep-alive re-proves access.
Talks to
CChannels
QUEUE · Redis pub/sub
Chat events, published once for every server to hear.
Talks to
QUEUE · Redis pub/sub
Notification events on their way to open tabs.
Talks to
QUEUE · Redis pub/sub
Which conversation each person is reading, shared between servers.
Talks to
QUEUE · Redis pub/sub
Who is on which sheet, and last-seen updates.
Talks to
DPlatform
SERVICE · in memory
Per-server and deliberately not persisted: after a restart it refills from the next broadcasts.
Talks to
SERVICE · messages · mentions
Stores the message, works out the recipients and publishes the event once.
Talks to
SERVICE · skips readers
Notifies everyone except the sender and anyone already reading the conversation, respecting each person’s settings; mentions also go by email.
Talks to
SERVICE · permission-checked
Owns the access check and the roster of who is on each sheet.
Talks to
SERVICE · RFIs · submittals
Status rules decide who holds the ball. Only an official answer moves it; a change of holder notifies the new one.
Talks to
Connections
Two flows
An RFI changes hands
Step through it and the parts it runs through light up above.
By count
Notes
Sales CRM. Turns a cold lead list into deals: it fills in missing contacts, runs email and LinkedIn sequences, and sorts replies so only real leads reach a person.
Live · Formwork and its clients
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APeople
UI · Next.js 15 · React 19
Formwork’s team: lead lists, campaigns, sequences and the CRM, across every client, each client’s data kept apart.
Talks to
UI · calls · follow-ups
Where reps work the follow-up queue, approve drafted replies and log calls.
Talks to
UI · per client
Each client sees their own campaigns, leads, pipeline and campaign health.
Talks to
EXTERNAL · MCP client
Claude, connected with a personal token, working the same data through MCP.
Talks to
BApp
SERVICE · CSV · domain dedupe
Uploaded lists, cleaned and matched to companies by web domain. Uploading never marks anyone as contacted.
Talks to
SERVICE · waterfall per field
Fills each missing field (email, phone, title, location) by trying providers in priority order, with rate limits and caching.
Talks to
SERVICE · email · LinkedIn
A sequence filters a live list into the leads it will contact, then hands them to the sender for its channel.
Talks to
SERVICE · events · replies
Receives the senders’ events. Only these can mark a contact as reached, which keeps the numbers honest.
Talks to
AI · 100 tools by role
Tools grouped by area and composed per role: an admin sees all of them, other roles only their own. Tokens are stored hashed.
Talks to
SERVICE · companies · contacts
Warm contacts and their companies, with status, temperature, notes and a relationship score from 0 to 100.
Talks to
SERVICE · proposal → won / lost
Deals, from proposal to won or lost.
Talks to
SERVICE · drafted answers
Real leads queue here with a reply drafted. A person approves it before it goes.
Talks to
AI · rules → Claude
Cheap rules catch bounces, unsubscribes and warmup mail first. Claude sorts the rest into six kinds, and says unsure rather than guess.
Talks to
AI · Claude · Firecrawl
Researches a contact or their company and preps a sales meeting, and answers questions about the CRM.
Talks to
CJobs
JOB · cron
Works through queued enrichment in the background.
Talks to
JOB · cron
Drafts answers ahead of time for confident real leads, so the queue is ready when someone opens it.
Talks to
JOB · cron
Pulls campaign stats from Smartlead and sends replies scheduled for later.
Talks to
DData
STORE · Prisma
Every client’s leads, lists, campaigns and CRM, scoped by client.
Talks to
STORE · Postgres · embeddings
A separate database of contact embeddings for semantic search in the assistant.
Talks to
EOutside
EXTERNAL · Proxycurl · Findymail …
Proxycurl, Findymail, Apollo and Hunter, among others, tried in turn for each missing field.
Talks to
EXTERNAL · email
Sends the email sequences, and sends approved replies on the same thread.
Talks to
EXTERNAL · LinkedIn
Runs the LinkedIn sequences.
Talks to
EXTERNAL · mailboxes
Connected mailboxes, for replies that arrive outside a sequence.
Talks to
EXTERNAL · models
Claude for triage, drafting and research; OpenAI for embeddings.
Talks to
EXTERNAL · triage alerts
Triage posts real leads to the team’s channels.
Talks to
Connections
Inside: reply triage
Most replies to cold email are noise. Triage stores every reply first, lets cheap rules catch the obvious, has Claude sort the rest, and only ever lets a person send, with the CRM updated before the one step that can’t be undone.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AArrive
EXTERNAL · reply webhooks
Replies to the email sequences, including ones from a colleague of the person emailed.
Talks to
EXTERNAL · mailbox webhooks
Replies that land in connected mailboxes.
Talks to
STORE · stored first
Every reply is saved before it is classified, so nothing is lost and anything can be reclassified later.
Talks to
BSort
SERVICE · every known alias
Smartlead has used several names for each field over time; extraction tries all of them.
Talks to
SERVICE · warmup · bounces · vendors
Catch warmup mail, auto-replies, unsubscribes and bounces without a model call. Vendor domains an admin pins are a hard rule.
Talks to
AI · Claude Haiku · six kinds
Sorts what the rules can’t decide into a real lead, unsure, vendor mail, automated, spam or warmup noise, with a confidence and a one-line reason. A bid invitation from a platform counts as real money.
Talks to
SERVICE · low → unsure
A low-confidence answer becomes unsure, so a person sees it.
Talks to
SERVICE · everything to people
One setting sends every reply to the hot channel for a person to read.
Talks to
CRoute
EXTERNAL · real leads · unsure
Slack, for replies a person should see now.
Talks to
EXTERNAL · Slack
Vendor mail, kept apart.
Talks to
EXTERNAL · every reply
Every classification, whatever the category, as an audit trail.
Talks to
SERVICE · one per reply
Real leads queue for a drafted answer; creating the entry twice for the same reply is impossible.
Talks to
DAnswer
JOB · cron
Drafts ahead of time for confident real leads; everything else drafts when someone opens it.
Talks to
AI · reply · tags · temperature
One call writes the reply and suggests tags (only from the client’s existing ones, never invented), a CRM status and a temperature from cold to burning.
Talks to
UI · CRM first, send last
The reversible work (status, temperature, tags) is applied first; the send, which can’t be undone, comes last.
Talks to
SERVICE · colleague replies
Finds the right thread even when the person replying was never in the campaign, using the id the original webhook carried.
Talks to
SERVICE · Smartlead inbox
Sends now or at a set time on the same thread, logs it, and marks the contact as replied.
Talks to
Connections
Two flows
A lead list becomes a pipeline
Step through it and the parts it runs through light up above.
By count
Notes
Sound-design instrument. A synth with six ways to make sound that plays in a browser tab or inside your DAW. Hunt breeds new patches from the ones you keep, so you find sounds by ear.
Invite-only for now · demos public
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APlay
UI · Next.js 16 · React
The instrument in a browser. The public build keeps everything in the browser, with no accounts and nothing stored on a server.
Talks to
UI · WebMIDI · computer keys
Play from a controller or the computer keyboard, in a chosen key and scale. Every browser API is optional; missing ones fail quietly.
Talks to
EXTERNAL · VST3 · standalone
Ableton or any VST3 host, on Apple Silicon, loading the plugin build.
Talks to
EXTERNAL · MCP client
Claude can list engines and presets, make and change patches, and write chord progressions.
Talks to
BInterface
SERVICE · WAV · drag out
Renders a sound offline to WAV, or records it, then lets you drag the file onto a DAW track (in Chrome) or download it.
Talks to
UI · one UI for both
The same interface in the browser and inside the plugin: engines, macros, the XY pad, presets and Hunt.
Talks to
SERVICE · mutate · breed · taste
Six variations at a time, judged by ear. Keeping some nudges the next generation toward them, within limits, and the nudge fades unless you keep choosing it.
Talks to
SERVICE · from parameters
A patch’s name and mark are computed from its parameters, so the same sound gets the same name everywhere, with nothing stored.
Talks to
AI · 13 tools · JSON-RPC
A plain HTTP POST endpoint. The protocol layer knows nothing about music; every tool body is a tested function.
Talks to
CSound
STORE · 84 across the banks
Starting points for every engine, plus a percussion bank.
Talks to
SERVICE · TypeScript · reference
The synthesis itself, in plain TypeScript: no browser APIs and no allocation in the render loops.
Talks to
SERVICE · analog · FM · physical …
Analog, FM, physical modelling, granular, wavetable and sampler voices, under one voice manager.
Talks to
SERVICE · two chains
An effects chain and a character chain, available on every engine.
Talks to
DHosts
REALTIME · separate realm
Runs the DSP on the browser’s audio thread. It is bundled on its own and only hears what the page posts to it.
Talks to
SERVICE · hand port
The same DSP in C++, kept in step with the TypeScript.
Talks to
SERVICE · WebView UI inside
Wraps the C++ engine as a VST3 and serves the React UI from inside the binary. Parameters are only ever appended, so saved automation never moves.
Talks to
EBuild
SERVICE · passcode gate · sign-ups
The private build sits behind a passcode; the public one has none. One setting decides, not a fork.
Talks to
STORE · sign-ups
Email sign-ups from the landing page.
Talks to
JOB · 197 golden renders
Renders fresh goldens from the TypeScript and measures the C++ against them on every build.
Talks to
Connections
Inside: the engines and Hunt
Each engine is a description plus a voice; the interface, the plugin parameters and Hunt are all built from the description. Hunt breeds sounds from a single seed, so a whole session can be replayed from one number.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
ADescribe
SERVICE · one line per engine
Adding an engine is one line here plus its module. No app code changes.
Talks to
STORE · params · ranges · groups
Every parameter with its range, default, curve and group. Linear and exponential parameters are normalised differently so a step means the same thing to the ear.
Talks to
UI · from descriptors
The controls draw themselves from the descriptor.
Talks to
BVoices
SERVICE · polyphony
Owns polyphony so engines never think about it. Every voice is allocated up front, so playing a note allocates nothing.
Talks to
SERVICE · antialiased saw · ladder
Subtractive voice with an antialiased sawtooth, so the top end sounds clean rather than harsh, into a ladder filter.
Talks to
SERVICE · velocity → brightness
Frequency modulation where playing harder opens the timbre, like a real tine, and nothing you do can push a preset past the brightness it was voiced at.
SERVICE · membrane · wood · metal · glass
A bank of resonant modes shaped like a drum head, a wooden bar, a bell or glass. Bowed excitation is bounded so it can never build up forever.
Talks to
SERVICE · shared buffers
Grain clouds panned across the field, sharing one buffer across voices instead of megabytes per note, with tapered windows so grains never click.
Talks to
SERVICE · sine → metallic
Tables that run from pure to complex: sine, triangle, soft organic, square, saw, two vocal formants and metallic.
SERVICE · your samples
Every voice plays one shared sample; swapping a 30 MB recording is a single reference swap between audio blocks.
CSignal
SERVICE · every parameter
Every parameter passes through a smoother. Unsmoothed changes cause zipper noise, the most common reason a digital synth sounds cheap.
Talks to
SERVICE · chorus · delay · reverb
Where the signal becomes stereo. Each unit is skipped when its mix is silent.
Talks to
SERVICE · bounded ±1.2
Effects that make a listener ask what that was, allocation-free and guarded against denormals, and bounded so nothing can push the output past ±1.2.
Talks to
DHunt
SERVICE · one number
A deterministic generator: the same seed gives the same hunt on every machine, so a session can be shared as a single number.
Talks to
SERVICE · six distinct takes
Walks the current patch to six variations, retrying until they sound distinct from each other.
Talks to
SERVICE · two kept parents
Crosses two patches you kept, with a little jitter.
Talks to
SERVICE · bounded · blended · fading
Turns what you keep into a direction for the next generation: capped so it can’t pin a parameter, blended so one odd keep doesn’t define you, and fading unless you keep choosing it.
Talks to
SERVICE · “brighter”, “looser”
Names the one or two things that changed in a player’s words, and keeps the trail of moves that found a sound.
Talks to
SERVICE · from parameters
Computed from the parameter values, so the same sound always has the same name and mark.
Talks to
Connections
Two flows
Finding a sound
Step through it and the parts it runs through light up above.
By count
Notes
AI testing platform. Point it at a GitHub repo and it tests the app for you: riskiest journeys first, in a real browser, with what broke explained in plain English.
Live · in development
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AClients
UI · Next.js · live map
Projects, the system map, plans, runs with replays and failure analysis, and a chat panel.
Talks to
EXTERNAL · OAuth · App · webhooks
Sign-in, repo access through an installed App, webhooks on changes, and checks and comments back on pull requests.
Talks to
EXTERNAL · MCP client
Trigger runs, read the system map, plans and results, and search failures from Claude.
Talks to
BAPI
SERVICE · NextAuth · REST
Workspaces, projects, plans and runs, plus the webhook receivers for GitHub, Linear, Slack and billing.
Talks to
REALTIME · SSE · pub/sub
Streams indexing progress, run events and browser screenshots to the dashboard as they happen.
Talks to
SERVICE · budgets · context
Coordinates the agents through a run: token budgets per agent, and context picked by similarity rather than sent whole.
Talks to
AI · 7 tools
Run, map, plan, results, health, failure search and a code audit, over MCP.
Talks to
CQueues
QUEUE · BullMQ
Jobs to clone and index a repo.
Talks to
QUEUE · BullMQ
Jobs to run test cases in a browser.
Talks to
QUEUE · BullMQ
Jobs to analyse results and send reports.
Talks to
QUEUE · BullMQ
Recurring runs on a schedule.
Talks to
DAgents
AI · system map
Turns a repo into a system map: framework, routes, components and their links.
Talks to
AI · journeys · cases
Rebuilds user journeys from the map, finds coverage gaps and writes test cases. Plans are cached.
Talks to
AI · runs · evidence
Drives a real browser step by step, classifies each failure, and captures screenshots, video and DOM snapshots.
Talks to
AI · plain English
Explains failures, spots patterns across runs, and files them where the team works.
Talks to
SERVICE · tree-sitter · regex fallback
Parses source with tree-sitter, falls back to patterns when it can’t, detects the framework, and indexes incrementally.
Talks to
SERVICE · 40 · 35 · 25
Scores each case on criticality (40%), risk (35%) and coverage gap (25%), and boosts journeys like sign-in and checkout.
Talks to
SERVICE · selector memory
Finds a moved element by stored fix, text, label or role before asking a model, and saves what worked.
Talks to
EData
STORE · pgvector
Projects, maps, plans, runs and healed selectors, with embeddings for retrieval.
Talks to
STORE · queues · pub/sub
Backs the job queues, rate limits and live events.
STORE · S3 / R2
Screenshots, videos and DOM snapshots from every run.
Talks to
FOutside
EXTERNAL · routed by task
Different models for mapping, planning, running and reporting, chosen per task.
Talks to
EXTERNAL · remote CDP
A real headless browser over CDP, or a cloud browser when configured.
Talks to
EXTERNAL · issues · alerts
Where failures become issues and alerts.
Talks to
Connections
Inside: the executor and self-healing
Real apps change under their tests. The executor checks it’s allowed to run, orders tests by what they depend on, and when a button has moved it climbs a ladder of cheap fixes before it ever asks a model, then remembers the answer.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
ABefore
SERVICE · never production
Refuses production environments outright, checks the target is reachable, and decrypts the test credentials.
Talks to
SERVICE · parallel groups
Sorts tests by their prerequisites into groups that run in parallel, one group after another, breaking any cycles first.
Talks to
SERVICE · skip repeat logins
Once signed in, later tests skip their opening login steps. Tests of the login itself always start fresh.
Talks to
SERVICE · plain HTTP
Contract tests run without a browser or a model loop, falling back to the browser when they need the UI.
Talks to
BRun
AI · one step at a time
Acts on the page step by step and classifies any failure: missing element, not clickable, timeout, wrong assertion, wrong page.
Talks to
EXTERNAL · Playwright · Stagehand
A real browser over CDP, with Stagehand’s element and prompt caching when it is used.
Talks to
SERVICE · stop, don’t pile up
If the AI provider fails on billing, auth or quota, the run stops rather than recording a false failure for every remaining test.
Talks to
CHeal
SERVICE · approved first
A fix that worked before for this selector, approved ones first, then others above 0.85 confidence.
Talks to
SERVICE · heuristic
Finds the element by the words on it.
Talks to
SERVICE · aria-label
Finds it by its accessible label.
Talks to
SERVICE · role-based
Finds it by what it is: a button, a link, a field.
Talks to
AI · selector sub-agent
Last resort: a model reads the page’s interactive elements and the selectors that failed, and proposes a new one.
Talks to
DAfter
STORE · saved fixes
Every heal is stored with its strategy and confidence, so the next run starts from it.
Talks to
SERVICE · quarantine
A test that fails then passes with no code change is flaky, not broken; after five such flips with under half its recent runs passing, it is quarantined.
Talks to
STORE · screens · API shapes
After a passing run, screenshots and API response shapes are kept to compare the next run against.
Talks to
STORE · screens · video · DOM
Everything a person needs to see what happened, kept per step.
Talks to
Connections
Two flows
Testing your app
Step through it and the parts it runs through light up above.
By count
Notes
Agent operations platform. A team of AI agents that works on a schedule. They read a company’s systems, draft the work and bring every decision to a person to approve.
Live · private
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AWeb
UI · Three.js room
The home screen: each agent a node with live presence, lines where one hands work to another, an artifact shelf and a chat dock.
Talks to
UI · j · k · a · r
One queue for every pending approval, gate and proposal, triaged from the keyboard, with a log of what was settled today.
Talks to
UI · artifacts
Every typed artifact an agent has produced, addressable by key.
REALTIME · socket.io
One shared connection per app; the Stage, chat and live thought streams all listen on it.
Talks to
EXTERNAL · fast-path chat
Talk to the agents from Slack, answered by the cheaper runtime.
Talks to
BEngine
SERVICE · node:http
The engine’s HTTP API, on Node’s own http module rather than a framework.
Talks to
JOB · GitHub · Linear · health
Gatherers that collect what changed (code, tickets, business numbers, tool health, memory) before each tick.
Talks to
JOB · BullMQ · 12 lanes
Health hourly, the inbox every two hours, reviews each morning, weekly analysis, nightly research. Lanes that share resources never overlap, and a circuit breaker pauses an agent that keeps failing.
Talks to
AI · Opus · Agent SDK
The top-level agent. It reads the signals, opens and claims a task, and delegates to the specialists.
Talks to
AI · Haiku · Sonnet · Opus
Sentinel, Analyst, Scribe, Custodian, Oracle, Architect, Strategist and the rest, each with its own model, tools, budget and yes-or-no checks.
Talks to
AI · OpenRouter · capped
A second, cheaper runtime for the frequent lanes, with a hard budget per tick, a turn limit, and only the MCP servers that lane needs.
Talks to
SERVICE · five
Seeds a team, its schedule and its autonomy defaults in one action.
Talks to
SERVICE · approvals · autonomy
Approval records and the autonomy mode (supervised, assisted or full) that decides what is approved automatically. New agents are capped at 20 per workspace and three a day.
Talks to
SERVICE · one run, start to end
Checks the budget, loads the agent’s past experience, filters its tools, runs it with a live thought stream, retries on failure, then writes the run to memory.
Talks to
SERVICE · claim · settle
The work record. The engine claims a task before delegating and settles whatever is still open when a run ends, so the record is right even if an agent forgets.
Talks to
REALTIME · agent → agent
Each run’s output is classified and published, so a Sentinel alert lands in the Orchestrator’s triage and an Architect deploy sets off a Sentinel check.
Talks to
AI · ~60 built-in tools
Assembles each agent’s tools from the healthy servers, minus the tools that agent is blocked from.
Talks to
JOB · 03:00 nightly
Simulates five outputs, grades them on yes-or-no checks, mutates the prompt, simulates five more, and promotes only at five percent better. Every experiment is logged.
Talks to
JOB · binary evals
Scores every run and drafts a prompt revision when quality drops. A revision may change at most a fifth of the prompt, and after three failed revisions in a row a person is asked.
Talks to
SERVICE · tick · day · month
Token budgets per agent per tick, per day and per month, under a global cap.
Talks to
SERVICE · spec → MCP
Compiles a REST or GraphQL connector from its spec into an in-process MCP server at install time.
Talks to
CMemory
SERVICE · memory service
The memory, as its own service: written after every run, consolidated on a schedule, searched before the next. Drawn in full below.
Talks to
DData
STORE · Prisma · pgvector
Agents, runs, tasks, approvals, budgets and experiments, plus the memory’s embeddings.
Talks to
STORE · knowledge graph
The graph of what the agents have learned.
Talks to
STORE · BullMQ · pub/sub
The heartbeat’s queues and the event channels.
Talks to
EReach
EXTERNAL · models
Claude through the Agent SDK; other models through OpenRouter.
Talks to
EXTERNAL · in-process MCP
GitHub, Linear and Railway: code, tickets and deploys.
Talks to
EXTERNAL · HTTP MCP
C0, ContentSphere and OutreachSphere, read through their own MCP servers.
Talks to
EXTERNAL · REST · GraphQL
Any API added through the hub.
Talks to
Connections
Inside: the memory
Memory isn’t a log. Everything an agent produces is read for what it’s about, merged with what is already known, consolidated every six hours, faded when unused, and searched by meaning, keyword and connection when it’s needed again.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AIn
SERVICE · successful runs
What an agent produced, sent to memory after the run.
Talks to
SERVICE · every run
A record of each run, failures included, so an agent can learn from what it has done.
Talks to
SERVICE · files · URLs · YouTube
Outside material ingested on request.
Talks to
AI · MCP ingest
Anything an agent chooses to remember through its memory tools.
Talks to
SERVICE · world models
A seed document becomes a model: actors with interests, tensions between them, dynamics and assumptions.
Talks to
BWrite
AI · ontology-aware
Pulls out entities and relationships using the known types and their properties. Runs on a local model when one is available, Haiku otherwise.
Talks to
SERVICE · spelling · meaning
Matches each entity to one already known: by fuzzy name (bigram similarity of 0.75 or more), then by meaning.
Talks to
SERVICE · confidence · half-life
Confidence comes from the source; lifespan from the type: 7 days for a signal, 90 for a client, 365 for a principle. Use keeps a memory fresh.
Talks to
SERVICE · merge · never drop
Writes nodes idempotently with bounded concurrency. If extraction fails, the raw text is kept as an observation rather than lost.
Talks to
AI · 1536 dimensions
Embeds each node for search. In production a failed embedding is an error, never a placeholder that would quietly poison recall.
Talks to
SERVICE · auto-link
Links new nodes to related ones already in the graph.
Talks to
CStores
STORE · L0 raw · L1 distilled
Nineteen core kinds of memory in Neo4j (new kinds can appear), from signals and bugs to decisions and principles, each fact able to say when it became true and when it stopped.
Talks to
STORE · pgvector
The vectors behind semantic search, cleaned up when a node is archived so it can’t haunt results.
Talks to
DKeep
JOB · every 6 h
Clusters similar memories, checks the cluster really is similar (cosine 0.65 or more), and replaces it with one distilled L1 memory. High-value memories are never compressed.
Talks to
JOB · after compression
Reads 12 to 60 recent observations, decisions and experiences together and writes up to eight new insights or patterns.
Talks to
JOB · rolling
Keeps a short summary on each node from its properties and its recent connections.
Talks to
JOB · daily
Recomputes each memory’s freshness from its half-life and last use; below 0.05 it is archived. A node with no timestamp is left alone, never archived on a guess.
Talks to
ERecall
SERVICE · before each run
The agent’s own past runs, ranked by overlap with the task at hand, added to its prompt so it builds a view from what it has actually done.
Talks to
AI · hypothetical answer
For abstract questions, drafts a likely answer and searches with that instead. Direct lookups skip it.
Talks to
SERVICE · dense + keyword · RRF
Runs semantic and keyword search and fuses the two rankings.
Talks to
SERVICE · 2 hops
Expands each hit two steps through the graph, so related context comes along.
Talks to
SERVICE · time · type · truth
Drops archived and invalidated facts, and can answer as of a past date.
Talks to
AI · against the question
A model reorders the shortlist against the original question, not the expanded one.
Talks to
Connections
Two flows
How an agent remembers
Step through it and the parts it runs through light up above.
By count
Notes
Open-source Claude Code plugins. Free Claude Code plugins for contractors. Tell Claude about your company once, and it drafts estimates, bid packages, RFIs and daily logs in your brand.
Live · open source
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AInstall
EXTERNAL · git repo
Fork the repo, configure it once, and push it to a private repo. Teammates clone it and get the company’s skills with no setup.
Talks to
EXTERNAL · plugin marketplace
The marketplace file registers six plugins; each installs on its own.
Talks to
EXTERNAL · skill zips
The same skills as standalone packages, for the web app.
Talks to
BSetup
UI · ten sections
A conversational interview, one section at a time. Each answer is written into the template tokens of every skill that uses it; re-run one section and only its tokens change.
Talks to
STORE · local Markdown
The company’s answers, kept in local config. Capabilities are read at run time rather than substituted.
Talks to
SERVICE · colors · fonts · voice
Keeps every document on-brand: colors, type, logos and how the company sounds.
Talks to
CSkills
SERVICE · 13 skills
The preconstruction pipeline: ROM, plan takeoff, conceptual budget, scope check, sub bid packages, bid leveling, formal bid, contract and subcontract.
Talks to
SERVICE · 13 skills
Everyday paperwork: proposals, RFIs and RFI logs, submittals, daily logs, look-aheads, change orders, punch lists, incident reports.
Talks to
SERVICE · 4 skills
For subcontractors: quotes, change order requests, T&M tickets and pay applications.
SERVICE · document engine
Renders everything as HTML or DOCX in the company’s brand.
Talks to
AI · agent
Checks an estimate before documents are generated.
Talks to
AI · agent
Writes project-specific scope by CSI division.
Talks to
AI · agent
Builds one self-contained bid invitation per trade.
Talks to
JOB · PostToolUse hook
Runs after every file Claude writes. A finished HTML document that still has an unfilled {{TOKEN}} is blocked.
Talks to
DOutput
EXTERNAL · HTML · DOCX
Estimates, bids, contracts, logs and the rest, ready to send.
Talks to
JOB · Python checks
Checks manifests, inventory, resource paths and template tokens. Unconfigured templates fail; there is no switch to skip it.
Talks to
JOB · dist build
Builds the standalone zips, after validation passes and a person reviews them.
Talks to
Connections
Inside: the setup wizard and its safety gate
Setup rewrites a company’s own files, so it is run like a change to production: planned, shown as an exact diff, approved, backed up, applied, and checked byte for byte, with nothing ever reported as done that wasn’t.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AInterview
UI · one at a time
Company, services, brand colors, type, logos, voice, contact, ideal client, capabilities and estimating defaults, each with a default you can accept.
Talks to
UI · only its tokens
Re-run one section and only that section’s values change.
Talks to
BPlan
SERVICE · outside live files
Answers and full proposed files are staged away from the real ones and validated first.
Talks to
STORE · before hashes · contexts
Every file, its hash before, the exact text and context of each replacement, and the fields that change. No global find-and-replace.
Talks to
UI · bound to the plan
The exact diff is shown, and approval is tied to that plan; change the plan and it needs approving again.
Talks to
CApply
JOB · backups first
Backs up every target and verifies the backups before anything is touched.
Talks to
SERVICE · stale hash voids it
Checks each file still matches its hash from the plan. If it changed since approval, the approval no longer counts.
Talks to
SERVICE · per file, per context
Applies only the approved edits.
Talks to
SERVICE · after hashes · fields
Reopens every changed file and compares it to the approved result, byte for byte and field by field.
Talks to
DOutcome
SERVICE · never a false success
Any failure, interruption or missing backup stops here, with the backups kept.
Talks to
SERVICE · before · after · conflict
On restart each file is classed from its actual bytes; finished edits are never repeated, and nothing is rolled back over someone else’s changes.
Talks to
STORE · measured, not claimed
Reports only what was measured: sections, changed files, checkpoint and any tokens still unfilled.
Talks to
Connections
Two flows
An idea becomes a formal bid
Step through it and the parts it runs through light up above.
By count
Notes
Content system. Each client’s newsletters, LinkedIn posts and web copy, planned from their brand and drafted from prompts written for them. The client approves every piece.
Live · internal
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APeople
UI · Next.js 15
Formwork’s content team: brand, plans, pieces, approvals and the calendar, one client at a time.
Talks to
UI · per client
Clients read their pieces, comment and approve.
Talks to
EXTERNAL · MCP client
Claude, connected with a scoped key, working the same content through MCP.
Talks to
BApp
SERVICE · versioned
Each client’s positioning, pillars and voice, with every version kept.
Talks to
SERVICE · audiences · frameworks
Who the client is talking to, and the messaging frameworks for each audience.
Talks to
SERVICE · client calls
Calls with the client, kept as source material for topics.
Talks to
SERVICE · monthly
The month’s topics per channel, drawn from the strategy and the calls.
Talks to
AI · 112 tools · scoped keys
A fresh server per request, with only the tool groups the key’s scopes allow. Keys are stored hashed.
Talks to
SERVICE · client × content type
The instructions for every pairing of client and kind of content.
Talks to
AI · draft · rewrite
Drafts and rewrites pieces from the matrix, with the client’s context.
Talks to
SERVICE · drafts · versions
One piece per channel, with versioned drafts, comments and a status.
Talks to
SERVICE · team → client
Internal review, then the client’s. A piece can’t enter review without real draft copy.
Talks to
SERVICE · scheduled
Approved pieces, on the dates they go out.
Talks to
CAgents
AI · Claude Agent SDK
The in-app agent. It hands work to the right specialist.
Talks to
AI · strategy → campaigns
Specialists from strategy and writing to prompts and campaigns, each limited to the tools its job needs, plus one tool to hand work to another.
Talks to
DData
STORE · Prisma
Clients, strategy, plans, pieces, versions and comments.
Talks to
EOutside
EXTERNAL · models
The models behind drafting and the agents.
Talks to
EXTERNAL · deliverables API
Formwork’s operations system, where each piece is also a tracked deliverable.
Talks to
EXTERNAL · notifications
Tells the team when work needs them.
Talks to
Connections
Inside: specialist agents and the prompt matrix
The in-app agent isn’t one model with every tool. An orchestrator hands work to specialists that each see only their own tools, and every draft starts from instructions written for that client, that kind of content and that reader.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AAsk
UI · in the app
Where the team asks for plans, drafts and campaigns in plain words.
Talks to
EXTERNAL · MCP client
Claude outside the app, using the same tools through a scoped key.
Talks to
BLead
AI · streams · tool rounds
Streams its answer, runs tool calls in rounds up to a limit, and delegates to one of six specialists.
Talks to
STORE · editable
Each agent’s instructions live in the database, so they can be tuned without a deploy, with built-in defaults if the database is unavailable.
Talks to
CSpecialists
AI · brand · audiences
Brand strategy, audiences, messaging frameworks, pillars and client context.
Talks to
AI · pieces · drafts
Writes and manages pieces: LinkedIn posts, newsletters, case studies, articles.
Talks to
AI · plans · calendar · shoots
Content plans, the editorial calendar, photo and video shoots, assets, and updates from construction projects.
Talks to
AI · prompts · personas
Designs the prompt matrix and personas, and tunes generation quality.
Talks to
SERVICE · per specialist
Each specialist gets its own list of tools plus one to hand work to another; nothing else is offered to it.
Talks to
DMatrix
STORE · client · type · persona
Instructions written for one client and one kind of content, optionally for one persona. These win when they exist.
Talks to
STORE · fallback
The house prompts, used when a client has none of its own.
Talks to
AI · traced to its prompt
The generated draft records which prompt produced it, so a bad draft can be traced to its instructions.
Talks to
Connections
Two flows
A topic becomes a published piece
Step through it and the parts it runs through light up above.
By count
Notes
Internal operations system. Runs Formwork day to day: clients, projects, tasks and invoices in one place, synced with the books. After each meeting, Claude drafts the follow-up tasks.
Live · runs Formwork
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APortals
UI · portal
The whole firm: money, clients, people and objectives.
Talks to
UI · portals
Managers see their team; members see their own tasks and time.
UI · portal
Clients see their services, billing, content, outreach and tasks.
Talks to
UI · portal
A prospect’s own view of what Formwork offers them.
EXTERNAL · MCP client
Claude over MCP, with tools scoped to the person’s role.
Talks to
BApp
SERVICE · JWT · role routes
Checks the session cookie, routes each role to its portal, and catches a role that changed since sign-in.
Talks to
SERVICE · proposals
Prospects and proposals, until one converts to a client.
Talks to
SERVICE · services · health
Each client and the services they buy, with one health score from 0 to 100.
Talks to
SERVICE · time logs
The work for each service, with dependencies and time logged.
Talks to
SERVICE · payments
Invoices, payments and expenses, allocated to clients and services.
Talks to
AI · 250+ tools by role
Tools composed by role: a client sees about ten, an executive over two hundred.
Talks to
SERVICE · transcripts
Each meeting and its transcript, with a badge until its drafted tasks are approved.
Talks to
AI · Claude
Drafts follow-up tasks from a transcript, with the context around the meeting.
Talks to
AI · facts · review
Claims mined from meetings, embedded and merged into facts. New ones are candidates until a second model verifies them.
Talks to
SERVICE · confidence tiers
Matches each cash line from the books to an invoice, a subscription or a person, with a confidence score. Only the surest can be approved in bulk.
Talks to
AI · search · describe · run
Above a threshold it gets a category index and three meta-tools instead of every schema, and loads only the tools it means to call.
Talks to
CJobs
JOB · scheduled
Flags tasks coming due and invoices gone overdue.
Talks to
JOB · scheduled
Sends reports and scheduled emails.
DData
STORE · primary · Prisma
Where every read comes from.
Talks to
STORE · dual write
Writes for tasks, time and money also land here, so the rollups people already use stay right.
Talks to
EOutside
EXTERNAL · OAuth · webhooks
The books. Its webhooks bring each change across.
Talks to
EXTERNAL · bank feeds · webhooks
Live bank balances and transactions.
Talks to
EXTERNAL · Formwork’s own apps
OutreachSphere and ContentSphere, read for each client’s outreach and content.
Talks to
EXTERNAL · alerts
Where the due-date and overdue alerts land.
Talks to
EXTERNAL · models
The models behind drafting, the knowledge base and the agent.
Talks to
Connections
Inside: meetings, tasks and what Formwork knows
A meeting does two jobs. It becomes tasks people approve, and it teaches the firm something: claims are pulled out, matched to the people and companies they’re about, deduplicated, checked by a second model, and fed back into the next meeting about that client.
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
AIn
UI · per meeting
The meeting’s transcript, internal or with a client.
Talks to
SERVICE · people · services · goals
Who was there, the client and their active services, and the firm’s current objectives and projects, so tasks land in the right place.
Talks to
AI · what we already know
What the firm already knows about this client, injected before analysis. A knowledge-base failure never blocks the meeting.
Talks to
BAnalysis
AI · Claude · in parallel
Writes the meeting summary and the follow-up tasks at the same time, each with an assignee, a service, a project and a due date.
Talks to
UI · badge until approved
Drafted tasks wait for a person. Until they are approved the meeting shows a badge.
Talks to
EXTERNAL · summary · tasks
The summary and the drafted tasks are posted to Slack.
Talks to
CLearn
AI · Haiku · Sonnet if unsure
Picks out claims worth keeping about the industry, the client or the business, tagged from a fixed vocabulary of 22 topics. Internal meetings need more confidence to count.
Talks to
SERVICE · names · aliases
Maps “Logan”, “Logan Smith” and “chief estimator” to the same person, following merges; a new name starts as a candidate for review.
Talks to
SERVICE · hash · cosine 0.92
An exact restatement or a near-duplicate becomes a new citation on the existing fact instead of a new row.
Talks to
AI · second model
Verifies or rejects new facts. Nothing is verified on the way in; borderline ones stay for a person to spot-check.
Talks to
SERVICE · superseded
When a newer claim replaces an old one, the old one is marked superseded rather than deleted.
Talks to
DRecall
STORE · pgvector · 1024-d
Canonical facts with their citations, embedded with one locked model whose name is stored beside each vector for future migrations.
Talks to
STORE · typed edges
People, companies and projects, linked to the account they belong to.
Talks to
SERVICE · verified-only
Semantic search with metadata filters, verified facts only by default so agents don’t retrieve noise, plus a keyed account spine, newest first.
Talks to
AI · MCP
Search, expand an entity, add or verify a fact, and the review queue, for any agent.
Talks to
Connections
Two flows
A meeting becomes work
Step through it and the parts it runs through light up above.
By count
Notes
Custom enterprise platform. Delta Construction’s project system, which replaced Procore: RFIs, bidding, contracts, chat and Autodesk files. Architects and subs reply from a link.
Live · Delta Construction
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APeople
UI · web · phones · iPads
Delta’s project managers, engineers and field staff, in the app.
Talks to
UI · emailed links
People outside Delta. They answer RFIs and submit bids from a link, without an account.
Talks to
BGateway
SERVICE · Spring Boot · JWT
Checks who is asking and forwards each request with their identity. Link tokens are checked here too.
Talks to
REALTIME · WebSockets
Pushes chat and notifications to open apps as they happen.
Talks to
CProject
SERVICE · team · directory
Everything hangs off a project and the companies and people on it.
Talks to
SERVICE · ball in court
Questions to the design team, with the ball moving between companies as the RFI is answered.
SERVICE · revise & resubmit
Submittals go round in revisions until approved.
SERVICE · sheets · revisions
Plan sets split into sheets, with revisions side by side.
Talks to
SERVICE · real time
Project conversations, live.
Talks to
SERVICE · invites · leveling
A trade’s scope packaged for bidding, with the subs invited to it.
Talks to
SERVICE · SOV · alternates
Each sub’s bid, with a schedule of values and any alternates, levelled against the others.
Talks to
SERVICE · SOV from the award
The awarded bid’s schedule of values is carried onto a payable contract on the same project.
Talks to
SERVICE · in-app · email
Tells people when something they own moves.
Talks to
DData
STORE
The record of every project, RFI, bid and contract.
Talks to
STORE · S3-compatible
Drawings, attachments and photos.
Talks to
STORE · pub/sub
Carries chat and notification events to the sockets.
Talks to
ELinks
EXTERNAL · one-time import
Projects came across from Procore once. Nothing writes back.
Talks to
EXTERNAL · folders · files
Delta’s Autodesk folders and files, linked to the project.
Talks to
Connections
One flow
A bid package becomes a subcontract
Step through it and the parts it runs through light up above.
Notes
Bid pipeline & CRM. Shows Delta’s estimators what they’re bidding and when: each bid gets a slot with a name, dates and hours, beside the client account.
Live · Delta Construction
Architecture
Choose any box for what it does, what it’s built with and what it talks to. Solid lines are requests; dashed lines are webhooks, queues, pub/sub and schedules.
APeople
UI · web app
The people who log leads and look after client accounts.
Talks to
UI · their schedule
Each estimator sees what they’re bidding, and when.
Talks to
BPipeline
SERVICE · statuses you set
Every bid opportunity, moving through statuses the team defines.
Talks to
SERVICE · calls · emails
Calls and emails due on a lead, logged when done.
Talks to
SERVICE · who · dates · hours
A bid on the estimating schedule: which estimator, which dates, hours planned and hours actually spent.
Talks to
SERVICE · specialties
Each estimator and their specialties.
Talks to
SERVICE
The companies Delta bids for, with their own statuses.
Talks to
SERVICE
The people at each account.
Talks to
SERVICE · timeline
Every call, email and meeting with an account, on one timeline.
Talks to
SERVICE · when it’s won
A won lead becomes a project.
Talks to
CData
STORE
The record of leads, accounts and the schedule.
Talks to
Connections
One flow
A lead gets an estimator
Step through it and the parts it runs through light up above.
Notes
Virginia, USA
I’m a systems architect in Virginia. In 2024 I co-founded Formwork, and I build the software it and its clients run on.
Since 2022 I’ve also been at CGI Federal, building AI for Medicare and Medicaid. On my own time I build a synth and AI tools that test software and run agents. I studied computer information systems at James Madison.
Nov 2024 to now
Technical Co-Founder & CTO
Formwork does sales, marketing and software for commercial construction contractors. I built all of its technology, and I decide how it fits together across the three units. It started as Sunrise Systems.
Jun 2022 to now
Senior Consultant1
I work in the Health and Social Services unit. I built the AI prototypes that led to a contract award from the Centers for Medicare & Medicaid Services. I also built the tracking systems and the dashboards leadership uses to see the team’s work.
1“He has been integral in the development of tracking systems and dashboards, automating and displaying team function at an executive level.”
Class of 2023
B.B.A. Computer Information Systems2
2“You have shown a speed and ability to solve problems that I have not seen in my students before.”
Off the clock
When I’m not at work I’m usually still building: SIGIL, Orbit and Thronos all started as side projects.
I read, I look after my two dogs, Pixel and Melon, and I follow racing, which is why this site is a race.
Guides for Formwork, letters on HEY World
Guides
I take on a few product builds outside Formwork. I design the system and write the code.
Tell me what you’re making. I read every message myself and reply within a couple of days.
2 product slots open for Q4
Form
