Kevin Fairley

Open to AI Research Engineer / Applied AI roles · LA or remote

Hi, I'm Kevin Fairley.

AI Researcher — Agentic Systems, Evaluation & Reliability

I make LLM agents measurable and dependable: tool-calling agents, multi-agent systems, eval harnesses, regression gates, and guardrails enforced in code, not prompts. Published with reproducible results and honest failure analyses.

  • Los Angeles, CA
  • U.S. Marine Corps veteran
  • English & Spanish

Selected work

Research & projects

All open source on GitHub. Live demos run without an API key and clearly label any demo or synthetic data.

MCP serverLive on Workers

LeadFlow MCP

An AI front desk for a small service business

A Model Context Protocol server that lets Claude, Cursor, or any MCP client run a shop’s lead pipeline: 11 tools, 3 resources, and 2 prompts to capture leads, score them with a published rubric, log conversations, schedule follow-ups, and report on the pipeline. It speaks stdio and Streamable HTTP on SQLite, and runs live on Cloudflare Workers with D1 behind a bearer token. Tested like production code: 113 unit and integration tests plus 4 end-to-end tests, 99.61% line and 96.88% branch coverage, and an 82% Stryker mutation score.

  • TypeScript
  • MCP SDK
  • SQLite
  • Cloudflare Workers
  • D1
  • Vitest
  • Stryker
Multi-agentLive demo

agent-crew

Multi-agent research desk

Planner, researcher, writer, and critic agents share a blackboard; you watch each handoff stream in live. The controls live outside the model: a deterministic check rejects citations to sources that were never retrieved, the critic loop is capped, and a token/dollar budget is checked before and after every model call.

  • TypeScript
  • Cloudflare Workers
  • React
  • Hono
  • SSE
EvalsLive demo

llm-watch

LLM call ledger & eval regression gate

Logs every LLM call's tokens, latency, and cost to SQLite, then gates model swaps in CI with an eval suite (regex, JSON-schema, and LLM-judge scorers), a paired bootstrap 95% CI, and pass-to-fail flip detection. In the demo suite it blocks a cheaper model that causes six pass-to-fail flips.

  • Python
  • FastAPI
  • React
  • SQLite
Agents

mini-agent

Tool-calling agent with self-check & eval harness

A provider-agnostic agent loop (OpenAI-compatible: Groq, Gemini, OpenAI) with calculator, sandboxed Python, and web-search tools, a step budget, a required self-check, and rate-limit-aware retries. Scored 14/15 (93.3%) on a 15-task eval with automated checkers; the published failure analysis shows the self-check accepting an off-by-one answer, motivating a re-execution verifier.

  • Python
  • pytest
  • Groq
  • Gemini
RAGLive demo

Personal Knowledge Engine

Retrieval-augmented generation over your own files

Chunks uploaded PDFs and notes in the browser, embeds them with Gemini, and streams a Grok answer that cites only the passages it actually retrieved. API keys stay on the server; without keys, it falls back to a clearly labeled keyword-retrieval demo mode.

  • TypeScript
  • React
  • Vite
  • Cloudflare Workers
  • Gemini embeddings
  • Grok
Guardrails

hermes-phone-agent

Guardrails for an open-source agent's phone/SMS skills

Wraps Nous Research's Hermes Agent telephony skill with guardrails enforced in code, not prompts: an allowlist, confirmation, call/text/spend caps, quiet hours, and an append-only audit log. Providers are mock-by-default, so it's safe to run.

  • Python
  • Hermes Agent
  • Twilio / Bland / Vapi adapters
ML systems

fast-tokenizer

Byte-level BPE tokenizer

Rust tokenizer with Python bindings. Exact GPT-2 token parity with tiktoken on 905,180 tokens; about 2.5× tiktoken's single-thread speed from Python, and 155.7 MB/s with 4 threads.

  • Rust
  • Python
  • PyO3
  • rayon
Data & analyticsLive demo

U.S. Economic Indicators Dashboard

Daily-refreshed macro dashboard on FRED data

A Python pipeline pulls 11 FRED series every morning in GitHub Actions and derives the signals economists watch: a real-time Sahm-rule recession indicator, yield-curve inversion spells since 1976, SAAR GDP growth, and the real policy rate. Growth rates are matched on the calendar, so the cancelled October 2025 CPI and household-survey releases can't skew a year-over-year comparison. The phone-first site charts every series with recession shading and 1Y–Max ranges.

  • Python
  • pandas
  • pytest
  • GitHub Actions
  • Chart.js
  • FRED API

More shipped apps

  • Turf Preview

    Phone-first generative-AI app: upload a yard photo, pick a turf style, get a photorealistic install mockup with a before/after slider. Photos are not stored. Live app is access-PIN gated.

    JavaScript · Cloudflare Workers AI · FLUX.2

  • Light Preview

    Same pipeline for holiday lighting: a dusk mockup of a house with Christmas lights, before/after slider, Save, and Share. Live app is access-PIN gated.

    JavaScript · Cloudflare Workers AI · FLUX.2

  • Pulse CRM

    Single-owner CRM: contacts, pipeline, tasks, and reports. The demo runs the real CRM handlers in your browser on fictional sample data.

    JavaScript · Cloudflare Workers · D1

  • EquiFlow

    Stock-tape dashboard on a simulated, clearly labeled price feed: SMA-20, drift alerts, and an optional server-side Grok read. Not investment advice.

    TypeScript · Cloudflare Workers

About me

Reliable agents, measured honestly.

I'm an AI researcher in Los Angeles focused on the unglamorous half of agentic AI: making agents measurable and dependable. I build open-source tool-calling agents and multi-agent systems, then wrap them with eval harnesses, regression gates, and code-enforced guardrails, and I publish the results, including the failures.

I design and direct my projects and build them with AI coding agents: I set the requirements, test against real models and real data, and report limits alongside results.

Before AI, I served in the U.S. Marine Corps and worked in server management, logistics, and data workflows. That background shows up in how I build: clear controls, audit trails, and systems that fail safely.

Skills

My toolkit

AI / ML

  • LLM agents
  • Tool calling
  • MCP servers
  • Multi-agent orchestration
  • Evals & regression gates
  • LLM-as-judge
  • Guardrails
  • Audit logs
  • Token/cost budgets
  • BPE tokenization

Languages

  • Python
  • TypeScript
  • JavaScript
  • Rust
  • HTML/CSS
  • English & Spanish

Tools

  • FastAPI
  • React
  • Cloudflare Workers
  • SQLite
  • PyO3
  • pytest
  • Playwright
  • Docker
  • GitHub Actions
  • Git

Experience & education

Where I've been

  1. 2024 – Present

    Data & Workflow Specialist

    Solvis Corp. · Commerce, CA

    Validate and integrate multi-source data, automate spreadsheet workflows, and build visualizations for leadership.

  2. 2015 – 2019

    Heavy Equipment Manager / Logistics Specialist

    U.S. Artificial Grass · Los Angeles, CA

    Managed inventory, freight, and equipment with computer-aided inventory, GPS, and diagnostic tools, including a military installation project.

  3. 2014 – 2017

    IT Specialist / Server Management

    Kesem Technology · Washington, DC

    Maintained server infrastructure for high availability and security compliance: patching, network health monitoring, and OSI-model root-cause troubleshooting.

  4. 2010 – 2014

    Heavy Equipment Operator, Lance Corporal (E-3)

    United States Marine Corps

    Maintained readiness of equipment, communications gear, and vehicles under strict security protocols.

Certifications & education

  • CompTIA A+ — earned 2014 · The Hope Project, Washington, DC
  • Job Readiness & Professional Development — Chrysalis, Los Angeles, CA · 2024
  • High School Diploma — James Irwin Charter, Colorado Springs, CO
  • Volunteer: Hospital Volunteer, Good Samaritan Hospital, Los Angeles, CA · 2025

Get in touch

Let's build reliable agents together.

I'm looking for AI Research Engineer / Applied AI roles in Los Angeles or remote. I'd love to hear from you.