I don't just prompt LLMs. I architect teams of AI agents that negotiate, veto each other, and keep a human in the loop — then I deploy them live.
The thing I care most about: the guarantee lives in the code, never in the model. A model can write the explanation; it does not get to be the reason you trust the number.
aurum-rails-uae-remittance — UAE → Global remittance settled on Arc in seconds, with every fee visible before you commit.
▶ Live app · 🎬 2-min demo · ⛓️ A real transfer this app made — settled in 3.6 s
Why this one, out of everything below:
- It refuses its own easy pitch. The obvious story is "remittances are expensive" — for the UAE corridor that is not true, so the README says so and competes on transparency instead.
- The FX spread is displayed even when it is 0.00% — showing the zero is what proves the line is real.
- There is a "what is real vs. simulated" table on the front page, listing what is not built (AED pay-in, KYC/AML) as prominently as what is.
- Money is
bigintminor units with branded types, not floats — because the sample it was forked from used floats, and that is not acceptable in a payment path.
Next.js 15 · Circle DCW · CCTP v2 · Arc Testnet · PostgreSQL
Agent systems
| Project | What it does |
|---|---|
| ⚖️ Contract Redline War Room | 5 agents redline contracts over Band. Compliance holds a veto that forces a visible re-plan; the human holds the only key. SHA-256 hash-chained audit trail you can verify from the UI. |
| 🔮 SupplyTwin | Supply-chain digital twin for UAE SMEs — predicts disruption ≥48h ahead. The 48h guarantee and the supplier-priority rule are deterministic code, re-checked at the API boundary; the LLM only supplies confidence and wording. Postgres RLS with FORCE, fails closed. 96 tests |
| 📡 Maya — CS Health Radar | Churn-risk agent that cites its evidence. Health score is deterministic TypeScript; a Postgres CHECK constraint makes an ungrounded recommendation impossible to save. Nothing is ever sent — the safety boundary is enforced by absence. Live |
| 📊 CompeteIQ | 4-agent GTM intelligence pipeline — full competitive briefs in under 20 seconds |
| 📄 DocuMind AI | 4-agent CrewAI document pipeline on AMD MI300X via Fireworks |
| 🎯 AMD Track 1 router | Token-efficient routing agent — Tier 0 solves what it can at zero tokens, and falls through rather than guess |
| 🛡️ PolicyForge | Plain English → enforceable security policy YAML |
| 💸 AgentFlow | Agent-to-agent USDC nanopayments on Arc L1 — sub-cent per task |
Tools I publish
| 🔌 3 MCP servers live on npm | x-growth-mcp (X · 7 tools) · linkedin-growth-mcp (LinkedIn · 3 tools) · threads-growth-mcp (Threads · 5 tools) |
Product & web
| 💼 Office CRM | Back office for a used-vehicle export business — lead → invoice → payment → unit → export documents. Next.js 16 · PostgreSQL |
| 🏘️ TokenEstate | Real estate tokenization — ERC-1155 fractional ownership, soulbound KYC, on-chain rent. 58 tests |
| 🤖 Physical AI & Humanoid Robotics | Interactive textbook with a RAG chatbot — grounded answers, cited sections, EN ↔ اردو |
Agent frameworks: OpenAI Agents SDK · AgentKit · ChatKit · LangGraph · LangChain · CrewAI · AutoGen · MCP (I build and publish servers)
AI/LLM: OpenAI SDK · Claude · Gemini · Groq · RAG · ChromaDB · Qdrant · Whisper · Prompt & Context Engineering
Practices: Spec-Driven Development — constitution → spec → plan → tasks → implement, with the artefacts committed
Backend: Python (FastAPI) · Node.js (Express, NestJS) · PostgreSQL · MongoDB · Redis
Frontend: Next.js 15/16 · React 19 · TypeScript · Tailwind CSS
Web3: Solidity · Hardhat · The Graph · IPFS · Wagmi · OpenZeppelin · Circle (DCW, CCTP, Arc)
DevOps: Docker · Kubernetes · Railway · Vercel · Supabase · Neon · GitHub Actions
"Shipping products, not just prompts."