
Computer Vision Engineer; reclassified as Principal/Consulting Software Engineer
IDVerse · Feb 2025 – Aug 2026
- Designed document-understanding pipelines across 10+ document types using OCR, layout/geometry signals, VLM/LLM inference, candidate resolution, schema-constrained outputs, and API integration.
- Prototyped CPU/browser OCR-KIE for a limited document subset, reaching roughly 5-second inference in a 40 MB Brotli-compressed WebAssembly bundle with experimental 60-70% accuracy and documented coverage limits.
- Fine-tuned a VLM for inference and an LLM for structured, domain-specific output over OCR covering 10+ languages; used LORA/QLORA, llama.cpp, schema constraints, and failure analysis.
- Built concurrency-aware FastAPI and Docker services using less than 3 GB RAM on CPU, with sub-minute inference and sub-30-second results in some cases; GPU inference approached 10 seconds.


