- 🤖 Building cited, fail-closed AI systems: retrieval with provenance, tool gates, and measured evals.
- ⛓️ Shipping blockchain and distributed-systems software across Kaia, Ethereum, and EVM-compatible chains.
- 🛡️ Focused on least-privilege tool use, reproducible evaluation, and agent safety that fails closed.
- 🌍 Working remotely at Kaia and building in public where I can.
Python TypeScript Rust LLM agents RAG & evals MCP Agent security Solidity Zero-knowledge proofs Distributed systems
- capitalcite — cited answers over U.S. bank 10-Ks. Every claim points at a filing span, retrieval changes are scored, and agent/MCP tool calls fail closed.
- s1-tool-gate — local per-tool-call allow / deny / escalate gate in front of a sandbox. Fail-closed deny, plus a pinned flip-rate and calibration check. Not a sandbox.
- gluecut — replace prompt/parse glue with a deterministic shell and typed decisions, scored against a schema-constrained LLM baseline.
- kaia-mcp — MCP server for Kaia accounts, tokens, NFTs, contracts, and calldata, with OAuth-scoped HTTP and fail-closed bearer auth.
- Conclave — database-free backend on a from-scratch Raft implementation: replication, snapshots, dynamic membership, linearizable reads, and a tamper-evident audit log.
- snarkjs-zk-demo — end-to-end ZK-SNARK demo with
snarkjsand a Solidity verifier.
Also: sft-loop, a measured PEFT LoRA and quantization loop on a pinned ≤3B model with no invented uplift, and swarm-world, a reproducible multi-agent simulation with a deterministic world and agent-free evaluation.
Contributor to T3 Code: merged a client fix that rejects incomplete GitHub owner paths before lookup (#23024), plus earlier provider-safety, Claude lifecycle, and provider-update UX fixes (#6602, #6606, #7761).
Contributor to LLM Wiki: Ollama web-search integration with API-key-free search.
I apply systems and protocol-engineering discipline to AI: answers that cite or refuse, tool calls that fail closed, evaluation you can re-run, and useful on-chain integrations.



