Case study · Equity research

Taufolio: equity research drafted from primary sources in minutes

My current independent build: an AI equity research platform that reads SEC filings and earnings-call transcripts, drafts a report with several models in parallel, reconciles them and anchors every claim to its source. A snapshot takes about twenty seconds, a full report about three minutes.

Own company Production

Mandate

Founded it and build it: the filing and transcript ingestion, the market-data layer, the multi-agent drafting and reconciliation, the citation anchoring and the report modes.

Key decisions

  1. Draft with several models in parallel — Claude, Gemini and GPT — then reconcile the drafts, rather than trusting a single model with the analysis.
  2. Anchor every claim to a citation in the underlying filing or transcript, so a reader can check the number rather than believe it.
  3. Two modes instead of one: a roughly twenty-second snapshot for a first look, a roughly three-minute full report when the question deserves it.

Outcome — measured

~3 min for a full report; about twenty seconds for a snapshot
6–8 h of manual research replaced per company Projected
Stakeholders
  • Me, as founder and the person building it
  • Investors doing their own company research, as the intended readers
Constraints
  • Primary sources are long, inconsistently formatted and filed on their own schedule
  • A wrong number in a research report is worse than no report
  • Market-data providers fail, rate-limit and disagree
  • One person building and running it
Reuse
  • Filing and transcript ingestion pipeline
  • Parallel multi-agent drafting with reconciliation
  • Citation anchoring
  • Quarter-over-quarter diffing

Context

Taufolio is the founder project I started in 2025 and my current independent build: an AI equity research platform that works from primary sources rather than from summaries of them.

Business problem

Understanding a listed company means reading its 10-K, its recent 10-Qs and 8-Ks and the last few earnings calls, then holding all of it in your head next to the price and the numbers. That is six to eight hours per company, which is why most people skip it and read someone else’s opinion instead.

My mandate

I founded it and I build it: ingestion, the market-data layer, the drafting and reconciliation, the citation anchoring and the report modes.

Decisions

  • Ingest the primary sources directly: SEC EDGAR filings — 10-K, 10-Q and 8-K — and earnings-call transcripts, parsed with Azure Document Intelligence so tables and structure survive the trip.
  • Take market data from Yahoo Finance with Finnhub failover, because a single provider outage should degrade a report, not stop it.
  • Draft in parallel with several models — Claude, Gemini and GPT — and reconcile their drafts against each other, so a single model’s confident mistake does not become the report.
  • Anchor every claim to its source with citation anchoring, so a reader can go straight to the sentence in the filing that a number came from.
  • Offer two depths: a snapshot in about twenty seconds for triage, and a full report in about three minutes when the company is worth the time.

Delivery

Beyond the single-company report, the platform does Ticker Compare for side-by-side companies, monthly portfolio recaps, and quarter-over-quarter diffs that show what changed in a company’s own words between filings.

Outcome

A full report in about three minutes, standing in for the six to eight hours of reading and note-taking the same coverage would otherwise take per company.

Reuse

The ingestion pipeline, the parallel drafting and reconciliation pattern and the citation anchoring are the parts that generalise beyond equity research to any domain where long primary documents have to become a defensible written answer.

Evidence

The report timings are measured on the running system. The six-to-eight-hour figure is my own projection of the manual research a full report replaces, not a customer measurement. The platform is at https://www.taufolio.com.

Want the same thing done in your environment?

This case is one of several. If the shape looks like your problem, the fastest route is to send me the constraints you cannot move.

Remote-first, on-site when it matters; NDA on request