All projects
Case study · 2024 — Present
SynthMind — Multimodal Document Reasoning
Long documents, tables, and structured extraction
Independent builder · Independent
Personal product
PythonPyTorchLangGraphLayoutLMOCR
The problem
The problem
Contracts and research PDFs break naive RAG: length, tables, multi-column layouts, and the need for structured answers.
Outcomes
- Hierarchical summarization + document graphs for 100+ page corpora
- Layout-aware segmentation improving table accuracy dramatically in internal evals
Approach
Segment layout before OCR/text extract; reason over retrieved sections, not whole files.
Vision + language + structured tools for financial/legal artifacts.
Key tradeoffs
Context strategy
Chose: Hierarchical retrieval
Rejected: Stuff entire PDF in context
Cost and attention collapse past a few dozen pages.
What was hard
- Layout diversity across scanners and exports
What I would change
- Public benchmarks with redacted samples
Artifacts
