Dashboard/PRISM/pipeline-orchestration-notes.md
PRISM

pipeline-orchestration-notes.md

PRISM Pipeline Orchestration Notes

This note consolidates the PRISM-specific material previously sitting inside WorkOS-WIP.

Origin

PRISM is the SaaS/product conversion of the algolia-search-audit skill system.

The original skill system automates prospect research and search audit for Algolia sales teams. It is production-grade but fragile because it relies on many isolated agents, file gates, manual shepherding, and strict evidence discipline.

Current PRISM v2 State

Known from the recovered WorkOS-WIP notes:

  • FastAPI + Temporal backend exists.
  • Next.js frontend with copilot UI exists.
  • 7 of 20 modules were registered in v2 at the time of capture.
  • Not feature-complete.

This should be re-verified before making current implementation decisions.

Pipeline Shape

Wave 1: parallel intel modules
  -> company
  -> tech stack
  -> traffic
  -> competitors
  -> financial profile
  -> investor intelligence
  -> hiring, social, news signals
  -> partner and industry intel

Wave 2: query generation

Layer 2: browser audit with Playwright

Layer 3: synthesis
  -> business case
  -> sales plays
  -> audit report and SPA
  -> ABX campaign assets

Layer 4: quality gate
  -> factcheck
  -> evaluation

Evidence Discipline

Every data point should carry provenance:

  • FACT: authoritative or directly fetched source.
  • ESTIMATE: clearly marked estimate with source and date.
  • WEBFETCH: page fetched directly.
  • WEBSEARCH: search query surfaced the source.
  • NO_SOURCE: drop from final output.

Hierarchy:

AUTHENTIC > WEBFETCH > WEBSEARCH > NO_SOURCE

Known Fragility

The old flow can take 4-6 hours because:

  • Agents run in isolated contexts.
  • Agents communicate through files.
  • Gates check file counts, byte sizes, and JSON validity.
  • Failed modules must be rerun.
  • Context exhaustion happens if too much research is loaded into one agent.
  • The browser audit needs real rendering, screenshots, and strict artifact generation.

Failures That Shaped The Design

  • Costco workspace overwrite: test/prod isolation must be explicit.
  • DSW SPA missing brand CSS: reports must use the render pipeline, not hand-written one-offs.
  • Context overload: split research across role/team boundaries and use context caches.

Coordinator To Build

The coordinator should replace manual shepherding.

It should:

  • Accept a domain or target account.
  • Dispatch modules in the correct wave order.
  • Run gates automatically.
  • Rerun failed modules when safe.
  • Store research data.
  • Generate AE-facing artifacts.
  • Provide interactive sales coaching.
  • Report status conversationally.

PRISM should not stop at audit output.

It should translate audit findings into:

  • Personalized discovery questions.
  • Objection handling strategy.
  • Deep-dive areas.
  • Sales pitch and positioning.
  • AE coaching.

Next Move

Before building more, verify the current PRISM repo state and produce a focused project overview using the My OS flow:

project overview -> ELT critique -> Arijit review -> workspace execution team