Web / AI
AI Document and Data Analysis Platform
Full-stack development, from document upload and search to analysis, source inspection and Gemma 4 integration.
Implementation
- Included adjacent paragraphs and same-document section and table references. Budgeted messages and schemas with a tokenizer, reviewed missing evidence selectively, and retained source passages and quotations through final synthesis.
- Split claims into batches that fit the input budget and requested verdicts by short ID. Rejected altered, duplicate or missing claims, restored original text, and calculated support across all claims.
- Fixed prose requests being mistaken for table requests. Required item-level explanations and evidence IDs, copied quotations directly from source text, and retained supported answers while marking unresolved items.
- Removed repeated token counting during table merging and reused page analysis within each request. Returned answers before evaluation and scheduled evaluation separately, yielding to active questions.
Outcome
Reduced median local processing from 2,969ms → 672ms with fixed responses and 20 runs per condition. Evaluated 25 claims in 5 batches in a live case and returned answers without waiting for Ragas evaluation.