I take a customer's problem to production. Then I prove it works.
Eight years in GTM, the last six months as a solo builder. Ten AI products designed, built, test-agented and deployed through local, staging and production. Every one of them was used by real people, starting with me. Some are paused now; the code, the tests and the screenshots below are from the repositories as they stand.
Local → Staging → Production
Every entry below ran that path. The right column shows the tests that gated each release, as they exist in the repo today.
Tooling Claude Code, autonomous sessions, test agents that run scripted passes against the live deploy before anything ships.
Security OIDC for CI, secrets in AWS, row-level security, per-user tokens. No shared API keys anywhere.
10
products shipped
2,800+
commits, own repos
800K+
lines of code
24
agents on a schedule
6
months
Ship log
Newest first. Dates and counts come from the repositories, not from memory.
Sep to Oct 2026An AI agent platform (seed stage, name withheld)Live; agents paused since Oct
The hub's home page, screenshotted live on 3 Nov 2026. Account names, fundraising figures and teammates blurred.
GTM Hub and Demo EngineCompany operating system with 24 scheduled agents
I led GTM and built the system that runs it. A hub with its own agent and a 190-file knowledge base: pipeline, 48 account plans, 14 battlecards, playbook, track record. Twenty-four agents ran on a schedule (inbox drafter, follow-through, investor desk, gap finder, PR tracker, commitments chaser, week planner) with hourly email and Slack checks. The demo engine builds a prospect-specific demo on top of 13 fictional-company bases I built (HR across four APAC countries, an MSP service desk, a restaurant group, a mortgage lender, payroll, leave), each with an assistant inside that acts as whoever is signed in.
Demo tester agent runs up to 14 scripted passes per build; 23 builds through the engine Selftest + smoke logs on every demo base Hourly progress, email and Slack checks still running; the scheduled agents have been paused since I moved on in October
Oct 2026Personal productLive, used daily
Bend and BuildHosted training and nutrition coach
A full coaching app I use every day. The in-app chat does everything the app does: logs food from a photo or a sentence, tracks every nutrient and flags the ones you keep missing, adjusts the program, reads the scale and the pantry from pictures. SMS coaching, an admin dashboard for support, and its own Claude connector so the coach can read and write the app's data.
Stack Cloudflare Workers, D1, Anthropic API (Sonnet for meal picks, Haiku for logging), Twilio SMS, MCP connector
Localwrangler dev
Stagingtest account
Productionon my phone
Separate test account for every live pass, never the real one Costs metered per model call Last deploy 2 Nov 2026
Aug to Oct 2026Own ventureLaunched with billing
On RecordThe AI content strategist that learns you like a $5K ghostwriter
Every AI LinkedIn tool is a writing tool pretending to be a strategist. On Record is a strategist that happens to write: live voice interviews build a voice profile, content pillars drive a deliberate journey, and it measures outcomes instead of likes. Built on five primitives (voice, operator, sponsor, campaign, intelligence) so a solo founder, an exec with an EA, and an agency running 20 clients are configurations, not code branches. A free AI-phrase checker, search guides, llms.txt, and a one-line pause switch that archives every scheduled job without deleting anything.
4,500+ core tests, 634 web tests Canary with a hard stop, smoke generation, speed harness 995 commits, 383K lines, 1,826 files Rules enforced in the DB, not documented: triggers and RLS
Aug 2026Own projectShipped 5 Aug 2026
20 m and 250 m, rendered from the repository on 3 Nov 2026. The alert bar is the live NWS feed status.
The DiveA descent from the surface to Challenger Deep
The homepage is a scroll from the sea surface to the floor of the Mariana Trench. Every metre is clickable and opens that depth's chamber: pressure, temperature, the creatures that live there, the history, and the projects you can join or fund. Underneath it sits a platform: live NWS marine alerts, a live globe, trenches, an ocean history timeline, fisheries, a diver's path, surf, community uploads, careers and volunteering. 480 sourced facts behind a verification gate that refuses to ship an unsourced one.
Verification gate executed on every release 0 of 480 facts unsourced 268 commits, 92K lines, 491 files in 24 days
Aug 2026Own projectWorking, used on a real stack
The report on sample data: $288,000 of a $432,000 pipeline in a leak state, $55,547 modeled recoverable, every number linked to records.
Anchor ScanFinds revenue leaks in a GTM stack and prices them
Feed it CSV exports from the CRM, sequencer, calendar and call recorder. It finds the leaks (missed follow-ups, single-threaded deals, intent nobody answered), prices what each one costs with math a CFO can audit, and hands the team work queues for Monday. The report is chat-first: dispute any evidence row and every number recomputes. Say "our win rate is 25%" and the whole model re-prices with your number labelled as yours.
Stack Python standard library only, 17.7K lines · Attio API client · Apollo loader · Fathom transcript miner · self-contained HTML report
Localsample data
Stagingsynthetic leaks
Productionreal CRM export
Planted-leak validation on synthetic data 9 test suites Coverage view separates "not built yet" from "your export lacked it"
Jul 2026At BlazelBuilt in one week
Blazel Mobile and the Live Voice InterviewerFull mobile client plus a real-time voice agent
Twenty-one screens: queue, ideas, drafts, calendar, leads, team, programs, approvals, audit, workspace, capture. The headline is a voice agent that interviews you out loud, asks sharp follow-up questions, and turns the conversation into post angles in your voice. Deepgram listens, Claude thinks, Aura speaks, all bridged through my server so keys never touch the phone.
Stack Expo, React Native, Node bridge, Deepgram Voice Agent, Claude, EAS dev builds
LocalExpo Go
StagingEAS dev build
Productionon device
Every backend route verified before the device build 14K lines in 8 days Audio pipeline 16 kHz up, 24 kHz down, echo-gated
Jul 2026Own projectLive mode on real calendars
The Availability LayerA scheduling policy engine for people and AI agents
The judgment layer between calendar pipes and whoever needs time. An agent calls schedule.book and gets a booked meeting against a real calendar. A human gets their life defended: hard and soft blocks detected, soft ones moved with one tap, private events hidden. Never a grid, always an answer.
Stack TypeScript, pnpm workspace, vitest, pglite, FakePipe stubs for CI, Google Calendar for live mode
Localplayground
Stagingstubs in CI
ProductionLIVE mode
vitest across the workspace, typecheck, lint Clickable playground with zero credentials 13K lines, 38 commits, one day
May to Jul 2026Own ventureStaging and prod on AWS
anchorgtm.ai and the product deck, rendered from the repository. The app itself needs its backend, so it isn't pictured.
AnchorGTMThe AI-native CRM that runs your outbound
A full CRM with an AI layer that researches buyers, plans multi-channel outreach, drafts every touch in the rep's own captured voice, and coordinates the motion across email, LinkedIn and phone. Chat is the product: ask for 15 buyers at an account or a follow-up from today's call and it happens inline. Rep edits and replies feed back into voice profiles and ICP. A background worker has the day's plan, briefs and drafts ready before anyone sits down. I ran my own outbound on it every day.
1,475 commits, 287K lines, 1,084 files Security pass: cross-tenant delete, chat role gate, webhook auth Deploy rule: verify the running image digest after every promote
Sep to Oct 2026Own ventureTwo packs on one harness
ForemanAgents that act
A case (an order, a service call, a report) is a durable object that knows its own state, decides what happens next, does it or tells a named person exactly what to do, sets its own next wake-up, and learns from the outcome. Two industry packs run on the same loop: a building-supply desk and an HVAC desk, each over a simulated system of record. The owner's voice is enforced in code, texting hours are kept by the harness, and a lesson the agent draws on its own goes on trial as an A/B policy before it becomes the rule.
End-to-end suite gates every deploy (typecheck + e2e) Guardrails in code: prices, stock and dates never come from the model Outcomes measured per policy variant
2026GMI Cloud AgentBox hackathonDeployed agent
Capture screen and the generated report, from the hackathon build.
workflow-scoutRecord a workflow, get a build-vs-buy answer
Record your screen, talk, or describe a workflow in the browser. A multimodal model works out what you're doing, then a build route and a buy route run in parallel into a decision engine that returns recommended vendors, a pricing estimate, the full build-vs-buy TCO, today's manual cost, and net savings with live sensitivity sliders. Capture stays client-side; the container only sees sampled frames.
Strict-JSON model contract: a bad answer is a null, never a crash Refine loop re-runs the buy route once on low confidence
Jul 2025 to Sep 2026At Blazel, founding GTMIn production
Deal-monitoring agentsRoughly 500K lines over three months
Agents watching every part of a deal: CRM state, DocuSign, LinkedIn activity, mid-deal upsell signals. The same system a Series C company I later interviewed with was hiring three engineers to build. Alongside it: $1.7M+ in closed revenue including a $1M+ strategic deal with eight stakeholders and a three-month paid pilot that converted.
Stack Internal, built with Claude Code on top of Attio, Apollo and Fathom
Localown data
Stagingteam pilot
Productionwhole team
Lesson that shaped everything after it: too expensive to maintain at that scale, which is why the next four products were built on harnesses
How I work
The same loop every time, whether the customer is a Fortune 500 HR team or me at 6am.
Sit with the problem
Discovery calls, transcripts, the customer's own system of record. The spec is written in plain English as a job description before any code: who the agent acts as, what it may touch, what it refuses.
Build on a harness
Guardrails live in code, not in the prompt. Prices, dates and permissions come from tools. Every action and refusal is logged against the record, with a trace of how the answer was made.
Test with agents, then ship
Scripted test passes run against the real deploy, not a mock. Local, staging, production, with a canary and a kill switch. What was live-tested and what wasn't gets said out loud.
Built to pass an audit
Security is part of the build, not a pass at the end. What that looks like in the repos above:
No long-lived keys
CI deploys through GitHub Actions with OIDC, so there is no stored AWS credential to leak. Secrets live in AWS Secrets Manager and Workers secrets, never in code, never in a chat. Every repo above ignores its env files.
The system enforces permissions, not the prompt
Agents act as the signed-in person with their own token, so the customer's system of record decides what they can see and do. Every action and refusal is logged against the record with a trace.
Tenant walls in the database
On Record enforces its rules with row-level security and triggers: a voice is owned permanently, nobody attributes a draft to another account, interview transcripts are owner-only. Anchor closed cross-tenant delete and unauthenticated webhooks in a dedicated security pass.
Keys never reach the edge
The voice interviewer bridges Deepgram and Claude through my server so no key ships in the mobile app. Demo platforms are deployed by a named person who enters the keys, with a rule that they are never pasted into a chat. Logging and admin access go over VPN.
A rule from the Anchor roadmap, 17 Jun 2026. The web task definition was pinned to a frozen image tag, so every web change that session silently never shipped while the worker, pinned correctly, did. Fix: repin to the digest, track :prod going forward, and verify the running image digest after every promote. Every pipeline above inherits that rule.
What I'm after
A seat where I sit with the customer, find the real constraint, build the system that fixes it, and stay accountable until the number moves. The titles vary.
AI strategist or deployment strategist at an applied AI company
Forward deployed, customer-facing builder
Founding GTM engineer or agentic GTM operator
Background before building: EMT, years in construction supply at L&W, restaurants, door-to-door. I know how a yard, a kitchen and a service desk actually run, which is where most agent projects go wrong. East Bay based, open to SF hybrid and travel.
