document-intelligence-agent
An agentic document intelligence system that understands documents, retrieves relevant evidence and answers questions with source-level provenance.
What I work on
-
Document AI
OCR and vision-language models for parsing documents, tables and layouts with Docling and Chandra.
-
LLMs and agents
Structured extraction, RAG, agentic workflows and MCP servers built for production use.
-
Evaluation
Measuring what actually works, from precision and recall to field-level extraction accuracy.
-
Inference and serving
Serving models with vLLM on autoscaled GPU infrastructure, with metrics in Prometheus and Grafana.
Projects
Open source
Docling: expose logprobs on the streaming API path
Added token log-probabilities to Docling's streaming VLM output, so confidence can be read while a document is still being parsed.
Writing
Why 80% accuracy can mean 0% found: accuracy, precision, recall and F1 explained
What each metric actually tells you after training a model, and why looking at only one of them can mislead you.
Beyond OCR: how Docling let me upgrade to a vision-language model without rewriting my pipeline
Why choosing a document parser for its output format, not just its OCR accuracy, made a model upgrade painless.
Tools I use
- AI and ML
- PyTorch, Hugging Face Transformers, BERT, LoRA and QLoRA, synthetic data
- LLMs and agents
- OpenAI API, LangGraph, structured outputs, prompt engineering
- Document AI
- Docling, Chandra OCR, table and layout understanding
- Infrastructure
- vLLM, LiteLLM, AWS EC2 Auto Scaling, Docker, Docker Swarm
- Observability
- Prometheus, Grafana, inference metrics
- Backend
- Python, Flask, Kafka, RabbitMQ, MongoDB, MySQL, Linux