CH3 /Software Engineer · Dec 2025 – Present
FIRSTHUMAN TECH.
Helps a company get answers out of thousands of its own documents, handwritten forms included, by building the search and reading systems underneath.
The job
Retrieval and document AI over 5,000+ enterprise documents, handwriting included.
The current main mission. 5,000+ mixed enterprise documents, some of them handwritten. The engineering problems: chunking that respects document structure, re-ranking to fix recall, an OCR path for semi-structured forms, and a Kafka ingestion pipeline where any step can fail and be replayed safely.
What I did
- 01
Built a retrieval system over 5,000+ enterprise documents using embeddings and vector search.
- 02
Shipped a handwriting OCR pipeline on Azure Document Intelligence for semi-structured enterprise forms.
- 03
Built FastAPI, PostgreSQL, Redis and Kafka services for concurrent document ingestion and reliable reprocessing.
The hard parts
- Raising retrieval quality across a large, mixed document set.
- Reading handwriting reliably on semi-structured forms.
- Keeping ingestion correct under concurrency, with safe reprocessing when a step fails.
Results
- Improved retrieval quality through chunking strategy and re-ranking.
- Handwritten enterprise forms can now flow into the same document pipeline.
