Skip to content
Abdelkarim Ain
Menu

Projects

applied ai · in progress · 2026

Evidence-First Document Intelligence

A retrieval system for trustworthy answers over real documents, with citations, evaluation, and explicit failure handling.

Year
2026
Status
wip
Role
Sole developer
  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • MLflow
  • Astro
  • Docker

Project brief

Replace this brief with the document domain, users, and source collection once the project exists.

Problem

What real document-search or question-answering workflow needs better retrieval, traceability, or multilingual support?

Retrieval pipeline

Document parsing, chunking, metadata, hybrid retrieval, reranking, answer generation, and citations.

Evaluation

Add a small labelled evaluation set and report retrieval recall, answer correctness, groundedness, latency, and cost.

Safety and limits

Document prompt-injection tests, unsupported questions, missing evidence, and how the interface communicates uncertainty.

Comments

Signed in with GitHub. Discussion is stored as a GitHub issue.