Indian equity investors spend hours per company per quarter reading concall transcripts, investor presentations, and annual reports. Most of the signal is buried in unstructured text. Story of a Stock turns those documents into structured, comparable, queryable research data.
Concallyser is the factory. Story of a Stock is the showroom. The contract between them is the Supabase schema.
The backend orchestrator (`run-seeds`) reads company codes from `seeds.txt` and iterates through four extraction phases per company. Each phase has its own JSON Schema, its own prompt, and its own normalizer.
Shared infrastructure: one unified prep entry point for CSV discovery, FY/quarter normalization, PDF caching, and Gemini Files API upload. A minimal LLM provider abstraction (Gemini default; Claude available for text-only flows). PDFs cached locally; Gemini file IDs cached with a 20-day TTL aligned to the provider's actual retention window — not picked from convenience.
The frontend is a Next.js app on Supabase, deployed on Vercel — a read-and-render layer over the same tables the pipeline writes. Because both halves only talk through the schema, the schema-first discipline on the backend is what keeps the frontend from silently breaking when a field shifts.
Metadata, business snapshot, industry analysis, transcript scoring, moat analysis.
Fact base of up to 12 forward-looking facts, top-3 ranked catalysts with also-considered rejections, and base/upside/downside scenarios.
Management guidance tracking with a credibility verdict from a fixed enum (high_trust → not_assessable).
Non-financial key-variable discovery and KPI history extraction.
Every phase has a JSON Schema; every model call uses structured JSON mode. The schemas turned out to be the codebase's most valuable asset — the Phase 5 also-considered structure (next-best catalysts with explicit rejection reasons) and the Phase 6 credibility enum are domain modeling that only comes out of painful iteration. If everything else were thrown out tomorrow, the prompts and the schemas would still port to whatever replaced them.
View deep diveNormalizers rank, filter, clamp, and de-duplicate model output before persistence. The model never produces the final row directly. Cheaper than stacking another LLM as a verifier — and actually deterministic about what's been checked.
View deep diveScores like Phase 5's `growth_score` are computed in Python after the LLM returns, not asked of the model. Reproducible across runs, auditable, tunable without re-prompting.
View deep dive