Skip to content

AI Engineer · US remote

I build production LLM systems, and the evals and guardrails that make them safe to ship.

Python and FastAPI services around LLMs: extraction, retrieval and tool use, with the eval gates and agent-security checks that catch problems before users do. Open to full-time AI Engineer roles.

See the work Resume GitHub
Terminal capture: uv run pytest reports 36 passed, then the replay receipt prints GOOD 1.00 passing and MUTATION 0.80 failing, and the approval sequence denies without a token, rejects model text, executes once and suppresses the duplicate retry.
Offline Python sample: a mutated label fails the eval gate, a write needs an approval token, and a duplicate retry is suppressed. Recorded 2026-10-05. View the full-size terminal capture. Clone and run it in 10 minutes.

01 · Ship it

Production LLM apps

A paid SMS lead-qualification deployment for Acuity Real Estate, and DocExtract's FastAPI, Claude and pgvector extraction pipeline.

02 · Measure it

Evals as CI

DocExtract re-scores committed extraction outputs on every relevant pull request, and the check fails below a floor. llm-reviewer-path shows the same pattern in a 10-minute sample.

03 · Secure it

Agent and MCP security

Pre-registered prompt-injection research showing a defensive prompt stops some payloads and misses others, which points to fixing tool dispatch instead (a recommendation, not yet measured at 70B). Plus an MCP toolkit with auth and a read-only SQL allowlist.

Featured work

Four repos, each one runnable

All work and run commands →
DocExtract fixture explorer showing stored invoice fields and labeled sample confidence.

Ship it · Measure it

DocExtract

Turns PDFs and scans into validated, searchable records, with an extraction eval that runs as a CI check.

95.5% weighted field-level score on a 28-fixture offline replay; the CI check fails below a 0.85 floor.

FastAPIClaudepgvectorRedis + ARQ
Case study → Code →
Diagram: untrusted tool output reaches inline XML dispatch and leads to canary exfiltration; substrate_auditor flags the substrate, and typed dispatch is the recommended fix.

Secure it

ai-redteam-notes

Pre-registered research on indirect prompt injection through agent tool output, turned into a CI-ready auditor.

With a content-trust prompt scaffold in place, one payload variant still exfiltrated a canary in 10 of 10 runs on Groq-hosted Llama-3.3-70B (one provider, n=10). The recommended fix is in tool dispatch rather than the prompt; it has not been measured at 70B.

PythonMCP harnessesPre-registrationCI gate
Case study → Code →
Library map: tool logic surrounded by auth, per-caller rate limits, cache, cost attribution and OpenTelemetry.

Secure it

mcp-server-toolkit

A library of opt-in decorators that wrap MCP tool logic with auth, a read-only SQL allowlist, caching, rate limits and tracing.

JWT/JWKS auth, a sqlglot read-only SQL allowlist and OpenTelemetry spans, with 627 collected tests.

PythonMCP SDKPyJWTsqlglot
Code →
Receipt: GOOD passes the eval gate, MUTATION fails it, an unapproved write is denied and a duplicate retry is suppressed.

Measure it

llm-reviewer-path

A 10-minute runnable code sample: eval gates, approval-bound writes and idempotent retries.

36 offline tests, no API key. A mutated label set fails the eval gate, and a duplicate retry is suppressed.

PythonpytestuvGitHub Actions
Code →
Paid client deployment Jan to Mar 2026

Acuity Real Estate

SMS lead qualification inside the client's GoHighLevel CRM, with Spanish detection and a bilingual human handoff. Client-reported 500+ inbound leads.

FastAPI and Redis, broker-defined qualification rules, and a kill-check that paused expansion if CRM writes or Spanish handoffs failed.

Read the case study

Contact

Open to full-time AI Engineer roles

US remote. Applied AI, AI backend and LLM platform teams.

Email me LinkedIn GitHub Resume