About the builder
I'm an agentic AI builder. I ship production software by orchestrating teams of AI agents under sprint plans and review gates — one person moving at the pace of a small team. And I pair it with data-driven insight — funnels formalized as state machines, drop-off root-cause analysis, source attribution — so the software gets pointed at the right problem.
Right now I apply both as an AI Builder at Cuemath — designing, shipping, and operating the production platform behind internal operations end to end. Outside work, the same playbook ships everything in Works.
AI-integrated development
The through-line of all the work above: one person shipping at team pace by treating AI as an engineering system, not an autocomplete. AI does the building. I own the outcome.
A git-guard hook makes touching main directly impossible — even for the AI.
I don't rely on the AI being right — I make being wrong expensive.
Fact-forcing gates
Before any shell command or new file, the agent must state what the command verifies, confirm no existing file serves the purpose, and quote my instruction verbatim. Kills assumption-driven mistakes.
Devil's advocate gate
An adversarial reviewer agent interrogates every plan against the real codebase before I approve it. If the plan can't survive the challenge, it doesn't ship.
Four-reviewer verification
Specialized TypeScript, React, database, and security review agents must pass before any branch is PR-ready. Every migration goes through the database reviewer; every auth or ingest surface through security.
Clarify, never assume
Open PRD questions are hard build gates. Agents are instructed to ask me rather than guess — no silent extrapolation, no plausible-looking wrong output.
Multi-agent orchestration
~20-agent build teams, each owning one work track with explicit file ownership. A master orchestrator sequences the sprints — only it opens PRs.
Context engineering
Every repo carries CLAUDE.md instructions, specs, and persistent memory — agents start each session already knowing the workflow and the history.
Tooling integrated, not bolted on
MCP connectors give agents direct hands: GitHub for PRs and reviews, Playwright for browser-driven E2E, trackers auto-synced from progress logs.
Verify in the running app
Nothing ships on "the tests pass." Changes are exercised live locally, on staging, and — for big milestones — in full end-to-end dry runs.
The net effect: AI handles volume and speed; the process handles correctness; I handle judgment, product decisions, and everything the AI isn't allowed to decide alone.
Data + Insights
Analytics earns its keep only when queries connect to decisions.
Funnel state modeling
Model acquisition, activation, and retention flows as state machines with explicit transitions. Makes the funnel legible to product, sales, and engineering at the same time.
Drop-off root-cause analysis
Pareto-first, then RCA on the top drop-offs. Identify the two or three transitions where 80% of loss happens and instrument them properly before optimizing anything else.
Source attribution
Track cost, quality, and long-term retention per source. The cheapest source at the top of the funnel is often the most expensive by month six — attribution surfaces the truth.
Decision briefs
Every analysis lands as a one-page brief with the finding, the recommendation, and the counterfactual. Analytics only earns its keep by unblocking decisions.
- Python·
- TypeScript·
- JavaScript·
- Swift·
- SQL·
- C++
- Next.js 14·
- React·
- Flask·
- Tailwind·
- Prisma
- PostgreSQL·
- Neon·
- Supabase·
- SQLite
- TensorFlow·
- PyTorch·
- Claude API·
- Gemini·
- Prompt eng.
- Claude Code·
- Multi-agent orchestration·
- Sprint-planned dev·
- Review gates·
- MCP (Model Context Protocol)·
- Prompt caching·
- Prompt evals
- Funnel state modeling·
- Drop-off RCA·
- Source attribution·
- Retention cohorts·
- Pareto
- GitHub Actions·
- Netlify·
- Docker·
- Branch rulesets·
- Playwright E2E
- Anthropic Claude 1012026
- Anthropic AI Fluency2026
- Oracle Cloud Infrastructure 2024 Generative AI Professional2024