Skip to content

Engine

How FINN works, and what it cost today.

Two pipelines. Deterministic code for every number; a small, cached model only where language needs reading.

Push

Filings in, a short brief out

  1. 01

    Ingest

    One market-wide NSE call per poll, filtered to your 18 and upserted on seq_id. A watermark with overlap self-heals gaps.

  2. 02

    Classify

    Rules on the subject first; GLM with a strict JSON schema only for what rules can't place, in batches of 10.

  3. 03

    Score

    Category weight × your weight × impact, clamped 0–100. Fraud or default is always Act.

  4. 04

    Enrich

    Results from the XBRL, people and deals from the PDF. Related filings collapse into one card.

  5. 05

    Digest

    Only ready cards, with source health — no 'caught up' unless every poll in the window succeeded.

Pull

Documents in, cited answers out

  1. 01

    Documents

    Annual reports and concall transcripts, fetched from an allowlist and checksummed.

  2. 02

    Chunk

    ~500-token passages split at headings, page ranges kept, tables whole.

  3. 03

    Embed

    EmbeddingGemma-2 on Cloudflare Workers AI, 768 dimensions, task-prefixed.

  4. 04

    Retrieve

    Full-text and vector search fused by rank, always filtered to the one company.

  5. 05

    Cite

    Each quote must appear in its passage and each number in the passage or the filed facts — or it's flagged.

Today's meter

live

Deliberately not done

  • Intraday prices — end-of-day bhavcopy only.
  • News, broker notes and model memory on the research side — primary documents only.
  • A composite 'credibility score' — the Promise Ledger reports a fraction with its denominator.
  • OCR for scanned PDFs — they're flagged, not silently skipped.
  • Multi-user accounts — the demo is open, with no sign-in.
  • The BSE feed — NSE covers the portfolio.