AI systems builder focused on LLM applications, coding-agent workflows, MCP integrations, and deterministic safety systems.
I build Python and TypeScript systems that use model reasoning where it helps and deterministic code where correctness, authorization, and failure handling matter. My work spans multi-model orchestration, model-output evaluation, agent tooling, data lineage, incident response, and proof-gated onchain workflows.
- Coding-agent workflows: Claude Code, Hermes Agent, OpenCode, Codex, Qwen Code, Cursor, Gemini CLI, and Windsurf
- LLM applications: structured outputs, prompt design, multi-model routing, embeddings and reranking, live-data grounding, and reasoning traces
- Evaluation and reliability: deterministic scoring, fail-closed gates, edge-case and failure-mode analysis, idempotency, and tamper/replay verification
- Agent infrastructure: MCP servers and clients, async FastAPI/Fastify services, GitHub integrations, CI/CD, and cloud deployments
Five-model Qwen/FastAPI incident-response system with live Sepolia grounding, reasoning traces, a hard human-approval gate, and 31 passing backend tests. Onchain execution is explicitly simulated; the orchestration, grounding, and safety gates are real.
LLM-assisted data-model PR review backed by DataHub MCP lineage. Deterministic code owns the SAFE / RISKY / BREAKING verdict, with 27 passing tests, CodeQL, CI, and a live Cloudflare/Neon deployment.
Proof-gated payment rail for AI-agent work. A deterministic verifier and confirmed Casper testnet anchor must succeed before payment can become payable. The project includes a real MCP server, 770 backend tests, and 23/23 contract tests.
Extended a fork with multiple contributors that extracts and normalizes coding-agent sessions, including messages, code context, diffs, tool results, timestamps, metadata, and model usage, for downstream evaluation and training workflows.
- Hermes Agent Guide: practical setup guidance for Nous Research Hermes Agent across Linux, macOS, WSL, native Windows, and VPS deployments.
- DataHub Skills: reusable agent skills for DataHub search, lineage, quality, and connectors.
Building verification-first AI systems: agents that can reason and act, with explicit trust boundaries, observable decisions, deterministic safety checks, and honest failure paths.