# Earnings Call One — Product Vision

## The product

Earnings Call One is a focused research product for reading and understanding
earnings-call transcripts. Its first job is intentionally simple: display every
available company transcript from the connected dataset in a fast, searchable
website deployed on Vercel.

The product should do one thing well before expanding. It should not inherit the
full PandaInvest feature set or become a general investing dashboard.

## The user problem

Earnings calls contain useful operating detail, but they are long and often
mixed with prepared language, repetition, selective framing, and answers that do
not directly address the question. Investors need a quicker way to find the
important evidence without losing the original words or context.

## Initial experience

1. Browse all available transcripts by company, ticker, quarter, and call date.
2. Search the catalogue and open a complete transcript.
3. Read a concise AI-assisted brief beside the source transcript.
4. Jump from every summary claim to the exact supporting passage.

## AI analysis principles

The AI layer is a research aid, not a truth oracle. It should:

- separate reported facts, management claims, guidance, and analyst questions;
- highlight material changes, risks, uncertainty, and evasive answers;
- compare management's statements with prior calls when prior calls exist;
- remove greetings, boilerplate, and repetition without hiding source context;
- cite the exact transcript passages behind every analytical claim;
- say when evidence is insufficient instead of filling gaps.

The product should not label management as truthful or dishonest from tone
alone. Credibility must be grounded in checkable contradictions, changed
guidance, unanswered questions, or later outcomes.

## Data and provenance

The initial corpus is the public
`kurry/sp500_earnings_transcripts` Hugging Face dataset, pinned to immutable
revision `f3ded372da8d18dc6ad98955c4558e34b5fe6d45`. Searchable metadata lives
in Supabase Postgres; complete transcript text stays in private Supabase
Storage and is hydrated only through the field-limited server endpoint. Each
record preserves:

- dataset repository and immutable revision;
- original record identifier;
- ticker and company name;
- fiscal year and quarter;
- call/publication date;
- source URL when available;
- original transcript text;
- a content hash for deduplication and auditability.

The dataset card labels the corpus MIT and limits its stated use to research and
education. That metadata does not independently prove ownership of every
underlying transcript, so any future commercial redistribution should receive a
separate rights review. The current reader keeps the archive private at rest,
attributes the pinned dataset, and links each record back to its source.

## MVP boundary

Included:

- responsive transcript catalogue;
- transcript reader;
- search and basic filters;
- Clerk sign-up/sign-in foundation;
- Stripe checkout/session/webhook verification foundation;
- Supabase schema and read API;
- a controlled import path for the Hugging Face corpus.

Not included yet:

- portfolio tracking, stock screening, feeds, or SEC workflows;
- predictive investment scores;
- autonomous buy/sell recommendations;
- billing enforcement before the reading experience proves useful;
- AI summaries that are not grounded in stored source passages.

## Success test

The MVP succeeds when a user can quickly find a company call, understand the
few material points, and verify every important AI observation against the
original transcript.