I build tools for voice interfaces, agent workflows, and retrieval systems. My focus is making their behavior easier to inspect: where the evidence came from, what changed between runs, and what happens when something fails.
Explore the repositories · Voice projects · Agent and RAG tools
|
Know when to listen. Voice turn boundaries, bounded audio buffering, and interruption-safe playback generations. Inspect the timing in a local timeline. |
Measure the conversation. Transcript error alignment, latency breakdowns, and regression comparisons. Turn speech-pipeline changes into inspectable results. |
Connect voice to evidence. Timestamped transcripts, local document retrieval, and concise answers with source excerpts. Optional transcription through faster-whisper. |
| Project | The question it helps answer |
|---|---|
| ChronoRAG | What was true at a particular time, and do the retrieved sources agree? |
| Agent Replay Lab | Did the recorded agent run respect its tool, approval, evidence, and budget rules? |
| Context Ledger | Which memories are still valid, where did they come from, and what fits in context? |
| Tool Contract Lab | Will changed tool schemas break existing callers? |
| RAG Budget Lab | How much retrieval quality do we retain at a smaller context budget? |
| Project | What you can run |
|---|---|
| Voice Command Router | Train a local intent classifier, inspect abstentions, and transcribe audio with an optional speech model. |
| Document Vision RAG | Read document images with Tesseract and retrieve evidence with bounding-box citations. |
| Project | What you can explore |
|---|---|
| Graph Evidence RAG | Directed evidence paths, lexical retrieval, and exact source citations. |
| Crossmodal Eval Kit | Image/text embedding rankings, multi-positive recall, and paired comparisons. |
| Model Route Lab | Quality, cost, and latency tradeoffs using recorded model outcomes. |
| Schema Fuzz Workbench | Structured-output validation, targeted mutations, and reduced failure cases. |
| Agent Handoff Kit | Durable task dependencies, expiring worker leases, and retry fencing. |
| Semantic Cache Workbench | Scoped vector caching, context fingerprints, expiry, and threshold evaluation. |
| Retrieval Firewall Lab | Inspectable heuristic admission rules for untrusted retrieval context. |
| Context Redactor | Scoped pseudonyms, sensitive spans, and JSON-path redaction receipts. |
Start with a runnable example. Each toolkit includes an offline path so its core behavior can be explored without a paid API account.
Make the decisions visible. Inspectable traces, source identifiers, explicit contracts, and reports help explain why a result was produced.
Test the failure paths. Boundary conditions, malformed input, stale state, and regressions belong beside the happy-path examples.
Document the limits. Heuristic baselines, synthetic fixtures, and optional model integrations are identified in the project documentation.
Python · SQLite · JSON Schema · GitHub Actions · audio processing · retrieval evaluation
Interested in a particular failure case or integration? Open an issue in the relevant repository with a small reproducible example.