LLM Apps

Ledger

Structured extraction from 40-page financial PDFs with schema validation and a confidence score per field, replacing a nine-person review team.

Extraction is a schema problem

The trick is not prompting — it is refusing to accept output that does not validate. Every extraction is parsed into a Pydantic model; a failure triggers a repair pass with the validation error included in the prompt.

Confidence that means something

Self-reported confidence from a model is close to worthless. What worked: extract each field twice with different context windows, and flag disagreement. That correlates with real error at r = 0.79.

More systems

AI Agents

Atlas

An agentic retrieval system that decomposes a question, routes across five source types and returns an answer with paragraph-level citations.

Python
LangGraph
pgvector
Cohere Rerank

Computer Vision

Sentinel

Real-time defect detection running at 60 FPS on a Jetson Orin Nano — quantised, distilled and deployed to 140 production lines.

PyTorch
TensorRT
ONNX
OpenCV

NLP

Corpus

An open evaluation harness for RAG systems: golden sets, retrieval metrics, LLM-as-judge with calibration, and regression gates in CI.

Python
pytest
Ragas
DuckDB