Shawn Skelly
I build AI systems and the evals that check them
LLM pipelines Retrieval and evals Agent workflows Local models
I'm an AI engineer in western North Carolina, and most of what I build has a model in it somewhere: retrieval systems whose eval suites gate CI, LLM pipelines where a deterministic scorer makes the decision and the model has to justify any correction, and a backyard bird station that runs vision and audio classifiers on a local GPU. The largest is TALON, a land-analysis platform where natural-language search runs as tool calls over LiDAR and parcel data. I came up through quality engineering, so I write the tests first and I don't trust a model's output until I've measured it. Every project started with a real problem I wanted to solve.
How I Work
Featured Projects
🎯 evalharness
A retrieval eval harness built like a test framework: deterministic metrics, git-snapshotted baselines, CI gating, and an LLM that audits but never scores.
Active Sep 2026
🌍 TALON
Land acquisition intelligence platform. Conversational AI search, 3D LiDAR visualization, ML owner classification, and deep parcel analytics across western North Carolina.
Active Oct 2026
🔎 Salesforce Vector Knowledge
A Lightning Web Component that runs hybrid semantic + keyword search over Salesforce Knowledge with LLM-synthesized cited answers, at $0.0004-$0.0009 per search instead of Agentforce's $0.10/action floor.
Active Oct 2026
🔍 SF-Assistant
Conversational codebase assistant for Salesforce projects. Ask questions in plain English, get grounded answers with call graph context. Zero external dependencies.
Active Jul 2026
🔬 RepoAudit
Codebase analysis across six dimensions: architecture, security, AI maturity, docs, portfolio, diagrams. Deterministic core plus opt-in LLM validation.
Active Aug 2026
🌌 Blue Ridge Skyline
A single-file planner for stargazing nights in the North Carolina mountains: moon windows, hourly cloud forecasts, 38 curated spots and 116 parkway overlooks, and each spot's real skyline raycast from USGS terrain and lidar.
Active Oct 2026
⚖️ Docket Lens
A Python pipeline that downloads, classifies, and reports on every attached filing in an FCC comment docket, built while I was commenting in that same docket, with the bias that put into it measured and corrected.
Active Sep 2026
Recently Active
- 🌍 TALON Active Oct 2026
- 🌌 Blue Ridge Skyline Active Oct 2026
- 🛰️ GOES-19 Ground Station Active Oct 2026
- 🔮 PRISM Active Oct 2026
- 🔎 Salesforce Vector Knowledge Active Oct 2026
Writing
Some of this goes back to where the instincts came from, before any of it was code. Some works out how I earn the right to trust what a dataset or a model is telling me. The rest is field notes from the work itself. Different starting points, one throughline: learning to read a system well enough to know when to trust it. Some ideas run long enough to need more than one part.
My Projects Started Leaning on Each Other
i kept telling my agents to go grab something from one repo and bring it into another. so i sent two more agents to map what had actually crossed, and it was not what it felt like, in either direction.
I Asked Three Rival Models to Kill My Spec
before writing a line of code for a new system, i ran the spec through four adversarial review passes: my own, then three frontier models from three vendors, cold context, sequentially. 25 recorded fixes later it froze with zero open decisions. what the process caught, what it got wrong, and why the misses were still worth money.
Let's talk
Hiring for an AI engineering role, or building an LLM system that needs evals it can pass honestly? Get in touch.