Developer working at the edge of AI, building small, specific tools.
I'm James. I build MCP servers, agents, and observability tools that sit between language models and real enterprise workflows. Insurance-industry roots, AI-native approach.
Focus
- MCP servers & agent toolingcore
- Deterministic preprocessingdeep
- Observability & synthetic monitoringapplied
- Multi-model routingdaily
Stack
- Python · FastAPI
- PHP · MySQL
- PostgreSQL · PostGIS
- Claude · GPT · Ollama · Qwen3
- Docker · nginx · n8n
Currently
- Shipping Travel Alerts MCP
- Iterating on multi-model routing
- Open to dev-role conversations
Elsewhere
Small, specific tools I'm actively building.
Day 1 honesty: these are active builds, not case studies yet. I'll add write-ups as each one earns them.
-
01Travel Alerts MCP
An MCP server that folds nine global travel-safety feeds — GDACS disasters, US State Department advisories, FAA NOTAMs, NOAA, UK FCDO, USGS, CDC, ReliefWeb — into two queryable tools for LLMs.
-
02NOTAM Decoder
A Python library that decodes 400+ ICAO Q-codes, expands ~300 aviation abbreviations, and parses both ICAO and FAA domestic NOTAM formats — so cryptic aviation notices reach the model already structured, not as string soup.
-
03SiteProbe
A crawler that hits target URLs through multiple simulated visitor profiles and flags divergent HTTP responses. Born from a real production incident where silent WAF rules were blocking legitimate customers.
-
04MockShield
A configurable mock target with a rule engine returning varied HTTP responses (200/403/429/503) based on request shape. Companion to SiteProbe — makes the rule-misconfiguration failure mode reproducible on demand.
-
05MyAssistant — local orchestrator
A FastAPI + Ollama assistant that routes requests across 16 tools and two Qwen3 models (a 0.6B router, a 4B reasoner) — with DAG-based context compression so long conversations keep full fidelity without token bloat.
-
06docredact
A local-first PDF redactor. Two-pass PII detection (Presidio + GLiNER zero-shot NER) wrapped around layout-aware extraction, with an interactive approve/deny review — so sensitive documents never touch a cloud model before redaction.
Notes from living inside LLM tooling.
Starting this portfolio honestly.
I'm James Gault, a developer and AI power user based in the US. My day job is running customer service operations at a travel insurance company — training, escalations, the edges of the product where real customers meet real systems. That vantage point is where most of my projects come from: tooling for problems I watched go unsolved.
I came up on PHP and MySQL in the late '90s and early 2000s, stepped away from development, and came back through AI. Today I pair with LLMs to build the things I used to build by hand — MCP servers, agents, observability tools — with the same instinct for the quiet plumbing between systems.
This portfolio is the start of something, not the end of one. No giant shipped-at-Stripe logos yet — just a point of view, a few live projects, and a commitment to write honestly about what I'm learning as I build.
If you're working on AI tooling, agent infrastructure, or insurance-adjacent dev work, I'd like to hear from you.
- Based
- United States · remote-friendly
- Open to
- Dev roles, collaborations, contract work, interesting conversations
- Reading
- Anything on agent evals, DX, tools for thought
- Writing
- Weekly-ish
If you're building something in AI, email me.
I read everything. I reply to most. Short is fine — a sentence about what you're working on and why is usually enough.