Work

Products and open source.

Real things that run. I lead with what they do and whether the tests pass, not vanity numbers.

Products

Shipped and in use.

Open source

Small, tested, runnable.

jsonmend

Python

jsonmend

A zero-dependency library and CLI that extracts and repairs JSON from messy LLM output: Markdown fences, prose, single and smart quotes, Python constants, unquoted keys, comments, trailing commas, and responses truncated mid-object.

CI passing
envaudit

TypeScript

envaudit

A zero-dependency CLI that audits environment variables across modern JS stacks: vars the code reads but nothing defines, dead vars, credentials committed to .env.example, and secret-looking names inlined into the client bundle.

CI passing
dynarq-shield

Python

dynarq-shield

A prompt-injection and data-leak firewall for LLM apps. Channel-gated detectors, Unicode normalization to defeat evasion, additive risk scoring, allow/flag/block. Pure stdlib, zero runtime deps.

CI passing
rag-pipeline

Python

rag-pipeline

A retrieval-augmented generation pipeline from scratch: chunk, embed, FAISS store, retrieve, and generate grounded answers with citations. Local no-GPU, optional hosted mode.

CI passing
dynarq-agent

Python

dynarq-agent

A minimal AI agent runtime with a tool and skill interface, a tool-calling loop, and pluggable backends. Six built-in tools with path-traversal and SSRF guards.

CI passing
llm-finetune-lab

Python

llm-finetune-lab

LoRA and PEFT instruction fine-tuning with prompt masking. TinyLlama by default, a distilgpt2 CPU smoke test, and a generate script to try the trained adapter.

CI passing
llm-eval-harness

Python

llm-eval-harness

Deterministic evaluation of model answers: exact match, contains, token F1, ROUGE-L, and retrieval metrics, plus an optional LLM judge and a report renderer.

CI passing