Projects
Key technical work and initiatives.
saucier
Reads two printings of one cookbook and returns a catalogue of sauces, every claim traceable to the line it came from. 151 preparations in the 1909 witness, 57 that state a parent, 94 that state no base at all — the measured bar anything cleverer has to beat. Deterministic extraction behind four layers with no runtime dependencies, so a model can be added later without touching the provenance guarantees. Corpus committed, so a clone runs offline. Counts are the v0.7.0 README census. MIT.
- Python
- PyPI
- Architecture
- CLI
- Open Source
vramfit
Fits large open models onto one GPU by measuring which layers survive being crushed. Scans per-layer quantization damage, solves a mixed-precision recipe against a hard VRAM budget, then packs it. Three published packs, each measured against its baseline with the losing numbers printed: Nemotron Super 49B on a 24 GiB card, Nemotron 3.5 Lightning 30B-A3B entirely on a 16 GiB card, and Gemma 4 31B solved for 86k tokens of context beside Google's own 4-bit build. Twenty recorded data points in the evidence ledger, losses included. MIT.
- Generative AI
- Quantization
- LLM
- Python
- PyPI
- Open Source
typevet
Asks an open model typed questions about text and images: yes or no, pick one label, or pick a rubric level. Each answer is read as a probability from the model's next-token distribution, never parsed from written text. Runs Gemma 4 31B on one 24 GiB card with llama.cpp, and the same code on vLLM. With a receipt photo attached it caught 6 of 6 swapped totals it had passed as text alone. 7 releases. MIT.
- Generative AI
- Multimodal
- LLM
- Python
- PyPI
judgevet
Typed Python client, CLI and MCP server for TypeSafe's Jev, a hosted model that answers yes-or-no, pick-one and rubric questions with a probability for each answer. The client behind the Jev calibration study on public scam and bank-support messages. 19 releases. MIT.
- Generative AI
- Evaluation
- Python
- CLI
- PyPI
adk-secure-sessions
The compliance gateway for Google ADK — encrypted session storage in 5 minutes. Drop-in replacement that encrypts state and conversation history at rest using Fernet, closing the encryption gap for PHI, PII, and financial data. 4 dependencies. Apache-2.0.
- Generative AI
- Google ADK
- Security
- Python
- PyPI
docvet
Python docstring quality vetting that catches what linters miss. 31 rules across presence, completeness, accuracy, rendering, and visibility — including git-based staleness detection via diff and blame. Production/Stable, ~3,300 downloads. MIT.
- Python
- CLI
- Code Quality
- PyPI
- Open Source
AI Agent Framework
Designed and built an enterprise generative AI agent framework adopted by multiple teams. Standardized how LLM-powered agents are built, tested, and deployed with built-in guardrails, observability, and prompt management.
- Generative AI
- LLM
- Python
- Architecture
OCR Document Pipeline
Architected a scalable document processing pipeline that handled 500K+ documents. Automated classification, data extraction, and validation, replacing a manual review process and significantly reducing turnaround time.
- OCR
- ML
- Python
- Cloud
Video Processing Pipeline
Built an automated video processing system for analysis and content extraction at enterprise scale. Designed the end-to-end architecture from ingestion through output delivery.
- Computer Vision
- Python
- Automation
CI/CD Workflow Design
Designed end-to-end CI/CD pipelines for ML model deployment across multiple environments. Established testing, versioning, and monitoring standards that became the team's baseline for all new projects.
- DevOps
- CI/CD
- MLOps
- Cloud
Want to collaborate?
Check out my GitHub for more projects, or get in touch.