Skip to content

PRD.md

Latest commit

 

History

History
89 lines (75 loc) · 5.28 KB

PRD.md

File metadata and controls

89 lines (75 loc) · 5.28 KB
Raw

Insight — Product Requirements Document

Problem

Researching a company today means opening eight tabs — a stock screener, SEC filings, Crunchbase, Glassdoor, Twitter, and three news sites — and manually synthesizing what you find. That's slow for anyone who needs a fast, structured read on a company: an investor sizing up a position, a founder scoping competitors, or a candidate deciding whether to take an offer.

Insight collapses that research loop into one query: type a company name, get a structured, source-grounded decision brief in under a minute.

Personas

Persona Job to be done What they need from the brief
Investor Decide whether a public company is worth researching further Valuation multiples, margin trends, analyst sentiment, competitive positioning
Founder Understand the competitive landscape before building or pitching Competitor overlap, market sizing, positioning, funding comps for private players
Job seeker Decide whether to take an interview or offer Growth signals, funding health, culture/sentiment signals, "what's happening now"

The same underlying data (financials, SWOT, news, sources) serves all three — the differences are in which section they scan first, not in what's collected.

What shipped (v1)

  • Dual-mode reports. Public companies get financial statements, ratios, analyst sentiment, and earnings. Private companies get a startup-native view instead (funding rounds, hiring signals, milestones, traction) rather than an empty financials section — the pipeline detects company type and branches the whole report structure, not just hides a chart.
  • Source-grounded generation. Every AI-written section (SWOT, positioning, summary, competitive landscape) is generated from live web search results fetched at query time, not model memory — with a confidence score and, as of this update, a per-section "based on N sources, most recent [date]" footnote so a reader can gauge freshness without leaving the page.
  • Market Voices. Analyst rating actions (upgrades/downgrades) blended with StockTwits retail sentiment — investor and general-sentiment signals side by side. Job seekers get a rough proxy for company momentum for free from data already collected for investors.
  • Research Library. Saved reports, favorites, and search history persisted locally — repeat lookups don't require re-querying paid APIs, and a user building a comparison set (e.g. a founder scoping five competitors) has a place to keep them.
  • Export to PDF. One button; the report strips chrome (nav, search, sidebar) and prints cleanly — the artifact is meant to be shared, not just viewed once.

What was deliberately cut (and why)

  • Social media embeds (Twitter/Instagram post previews). Reading platform APIs now requires paid tiers ($200+/mo for X) with no free public read access for arbitrary accounts, and scraping violates ToS. Cut rather than shipped as a brittle/gray-area feature. Revisit via search-snippet retrieval (Tavily indexes tweet/post content) + Twitter's free oEmbed endpoint for the one platform where a clean, ToS-compliant embed is actually possible.
  • Employer rating card (Glassdoor/Indeed score). Same constraint — neither platform exposes a public ratings API, and both actively block scraping. Cut for v1; the plan is search-snippet retrieval (rating appears in the search snippet itself, which is legitimate RAG, not scraping) blended with hiring-signal proxies already in the data (employee growth, layoff news) rather than depending on one fragile source.
  • Compare mode (two companies side by side). Real user need — investors and job seekers both naturally compare — but deferred until the single-company report is fully solid, since compare mode roughly doubles the surface area of every existing bug.

Success metrics (if this were shipped to real users)

  • Time to first insight: seconds from query submit to first meaningful content render (currently gated by the slowest upstream API call in the pipeline). Confidence-footnote and streaming partial results are the next lever here.
    • Note (as of 2026-07-01): first-call latency currently also depends on OpenRouter account credit balance — a request that exceeds the available balance now fails fast with a clear 502 instead of a silent CORS-masked "Failed to fetch," but the underlying fix for speed is unrelated to this incident and still open.
  • Repeat-query rate: % of searches that hit the Research Library cache instead of re-querying — a direct proxy for whether the tool earns a second visit.
  • Confidence-score correlation: whether reports with score >80 get saved/ favorited at a higher rate than <60 — validates whether the score is actually signal or just decoration.

Prioritization rationale

Everything in v1 is either (a) needed to make both company types (public/private) usable, or (b) cheap to add given data already being fetched for another section (Market Voices' retail sentiment is nearly free once analyst data is already in hand). Anything requiring a paid API tier or ToS-risky scraping was cut rather than shipped in a degraded form — a broken or gray-area integration is worse for trust than an honestly-missing section.