Projects
Everything I've built, with a page each for the details. Source is open unless marked private.
Product
-
Orbit
Orbit · build brief product · 2026 privateTurns customer feedback into ranked product decisions, with the evidence attached.
Synthetic demo: 300 feedback items → 20 problems → 5 briefs
Systems
-
Trace Autopsy systems · 2026
Finds why an agent run failed by reading the trace, with no model in the analysis path.
Case-study example: the same tool call repeats 8 times
-
Agent Firewall systems · 2026
Answers one question before a tool runs: allow, hold, or block.
First matching rule decides; unmatched calls are blocked
-
AuditLens systems · 2026
Checks a dataset for unfair patterns before a model is trained on it.
COMPAS demo: 29 findings across 7,214 rows; small groups limit interpretation
-
doc-verdict systems · 2026
Turns "update the docs" into a test that either passes or fails.
Two task benchmarks; dry runs test the gate, not a live model
-
QualModel systems · 2026
Turns open-ended survey answers into a model where every claim still carries the quote it came from.
Extracted claims retain their supporting survey quotes
-
EngageMind thesis · 2026
A study assistant that answers from your own material, and can be fine-tuned to one learner.
Thesis prototype; not hardened for multi-tenant use
Research
-
FailSafeRAG research · 2026
Detects when document search has gone wrong, and recovers instead of answering confidently from bad evidence.
Research implementation; default visual recovery uses a mock model
-
ACOM embedding grid research · 2025
Places documents on a discrete grid so the layout can be read, then measures what structure survived.
Compared with PCA, t-SNE, and UMAP using shared metrics
-
LLM bias evaluation research · 2025
Tests whether a model's answer changes depending on who it thinks is asking.
Exploratory: 426 responses from 1,152 attempted runs