Most AI product people can't read an agent trace. Most engineers can't ship a product. I've spent 6+ years doing both — building multi-agent systems from architecture to production, and delivering 0→1 products in regulated environments where correctness isn't optional.

My edge is knowing where agentic systems break: tool-call failures, memory drift, non-deterministic planning loops, context handoff between frameworks. That isn't something you pick up from papers. It comes from shipping — and from four years writing smart contracts, where a bug is permanent and public.

6+ years · Multi-agent infrastructure · Smart contracts · RWA tokenization · 0→1 delivery

I publish tests, not screenshots. Every library below runs a real suite — no mocked happy paths standing in for coverage.

271 tests passing across 11 published libraries.

dream-memory shipped to npm, then went through an external principal-engineer audit — 3 critical bugs and 2 security findings fixed, each with a regression test. The audit report is public in the repo. I'd rather show the review than pretend it went perfectly.

dream-memory — Persistent, reflective memory for AI agents. 4-phase consolidation loop, two-model retrieval, zero runtime dependencies. Published as @dream-memory/core. TypeScript · 28 tests · externally audited

agent-governance-kit — OWASP Top 10 for Agentic AI as deterministic middleware. Prompt-injection screening, PII redaction, tool allowlists, loop detection, human-escalation gates. Sub-millisecond overhead, no LLM in the enforcement path. Python · 38 tests

relay — Cost-optimizing LLM router. A deterministic sub-1ms classifier tiers prompts and routes cheap work to cheap models — while correctly keeping hard prompts on frontier models. Budget gates with a thread-safe spend log. Python · 55 tests

prism — Reversible semantic context compression. 56–60% token reduction on agent traces, lossless on URLs, emails, code, and numbers. Python · 31 tests

context-fabric — Typed context envelope for cross-framework agent state. Adapters for LangChain, AutoGen, CrewAI, and OpenAI. Fills the state-handoff gap MCP left open. Python · 19 tests

temporal-cache — Time-aware semantic LLM cache. A 4-tier staleness classifier means "bitcoin price" and "capital of France" don't get the same TTL. Python · 16 tests

agent-mesh — P2P multi-agent coordination via stigmergy. Agents leave signals in a shared store instead of messaging each other — no central coordinator, no orchestrator bottleneck. Python · 13 tests

vault-sync — End-to-end encrypted portable agent memory with CRDT sync. Your agent's memory moves across devices and providers without a server that can read it, and merges conflict-free. Python · 22 tests

skill-market — Verifiable AI skill registry with behavioral provenance. Signed attestations, so an agent can prove what it actually does rather than what its description claims. Python · 19 tests

meeting-less — Agent that converts meeting requests into async decision flows. Turns "can we hop on a call?" into a structured decision document. Python · 19 tests

nexora — Agent-native data layer. SQLite by default, Postgres when you need scale, exposed as 38 MCP tools. TypeScript

forge — AI-assisted ERC-3643 RWA tokenization scaffolder. Describe an asset in plain English, get a compliance-checked deployment config and contracts. Compliance rules are hardcoded, deliberately — an LLM hallucinating a securities rule is a legal problem, not a UX one. Python · Solidity · ERC-3643 / T-REX

ledger — On-chain tamper-proof audit trail for AI decisions. SHA-256 commitments on Polygon at ~$0.00003 per decision, mapped to EU AI Act Article 12. Python · Solidity 0.8.20 · Hardhat

mcp-blockchain — MCP server exposing on-chain data as native agent tools. Read-only by design: no signing, no key handling. An agent using it can observe chain state — never move funds. JavaScript · ethers.js · 11 tests

clui-cc-windows — Desktop overlay for Claude Code, ported to Windows. Multi-tab sessions, permission UI, transparent floating interface. TypeScript · Electron

Agent infrastructure & governance   ████████████░░  Shipping
MCP tooling & protocol layer        ██████████░░░░  Building
RWA + AI convergence                ████████░░░░░░  Shipping
Agent evaluation & observability    ██████░░░░░░░░  Researching

Listed by capability, not by model version — point releases rot faster than any README gets updated.

LangGraph · LangChain · CrewAI · AutoGen · OpenAI Agents SDK · Anthropic Agent SDK · Pydantic AI · Semantic Kernel

MCP (Model Context Protocol) · A2A · Function / tool calling · Computer use · Claude Code · Gemini CLI

Frontier and open-weight families across Anthropic Claude · OpenAI GPT · Google Gemini · Meta Llama · Mistral · DeepSeek · Qwen — selected per workload on cost, latency, and reasoning depth rather than brand loyalty. See relay for how I automate that tradeoff.

LlamaIndex · Weaviate · Pinecone · Chroma · LangSmith · OpenTelemetry · Phoenix

Vertex AI · Azure AI Foundry · AWS Bedrock · Groq · Together AI · Fireworks AI · Docker · PostgreSQL · Redis

Solidity · Hardhat · ethers.js · Web3.js · OpenZeppelin · Chainlink · ERC-3643 / T-REX · Hyperledger Fabric · R3 Corda

TypeScript · Python · Next.js · React · Node.js · JIRA · Linear · GitHub Actions · Agile

6+ years across blockchain engineering and AI product delivery.

• AI Product Manager & Web3 Program Lead — agentic infrastructure, multi-agent orchestration, 4+ parallel product lines
• Web3 Project Manager — shipped blockchain features into a live platform serving 100M+ users
• Blockchain Developer — DeFi protocols, NFT platforms, token vesting, multi-sig governance, security audits and gas optimization
• Enterprise blockchain — Corda supply-chain systems, multi-party transaction flows, tamper-proof audit trails

Certifications: R3 Certified Corda Developer · NPTEL Blockchain (IIT, Govt. of India) · Startup India Maharashtra finalist · IICDC, IIT Madras

Remote-first, global.