Chiara Greco

Financial-research assistant

Vector index over 10 years of analyst reports. Hybrid retrieval, Claude synthesis with mandatory citations.

Adopted by 6 analysts; saved 4h/week each.
0

Problem

Six equity analysts re-read past coverage every time a thesis shifted. Memory was unreliable, search returned PDFs not paragraphs, hours wasted weekly.

Users

Internal — 6 senior equity analysts at a boutique

Key flow

  1. 1Analyst asks question scoped to sector or company
  2. 2System returns synthesized answer with citations to source reports
  3. 3Analyst saves the thread or refines the question

What I built

  • Ingestion of 14k analyst reports with table + chart extraction
  • Hybrid dense+BM25 search with date and sector filters
  • Claude synthesis with mandatory inline citations — no uncited claims
  • Per-analyst saved-question library for follow-ups across weeks

Limits

  • No real-time market data — covers research history only
  • Citations point to PDF page, not paragraph (limitation of source format)
Financial-research assistant

AI tools used

  • LangChain
  • Together AI
  • Supabase