RAG / Product StrategyMock demo

SignalDesk AI

An AI product-intelligence workspace that ingests user feedback, clusters pain points, finds evidence, and generates PRDs, roadmap bets, and experiment plans, every claim cited.

Next.jsTypeScriptAnthropic APIOpenAI embeddingsSupabase pgvectorVercel
Live demoGitHubAll projects

At a glance

ICP
Early-stage SaaS teams and founders with 50–500 feedback items from CSVs, app reviews, or notes.
Features
  • Upload feedback CSV or paste notes
  • Embed + semantically cluster themes
  • Evidence-backed insights with citations
  • Generate a PRD from validated opportunities
  • Generate roadmap + experiment plans
  • Track product-metric assumptions

AI architecture

  1. 1
    Ingest
    CSV / pasted notes, validated and normalized.
  2. 2
    Chunk + embed
    OpenAI text-embedding-3-small (real) or deterministic local vectors (mock).
  3. 3
    Cluster
    Group feedback into themes; score opportunity size + confidence.
  4. 4
    Retrieve
    pgvector similarity search fetches top evidence snippets.
  5. 5
    Generate (RAG)
    claude-opus-4-8 drafts PRD / roadmap / experiments grounded in citations.
  6. 6
    Verify
    Guardrail requires ≥2 evidence links per recommendation.

Case study

Product problem

PMs need to move from a pile of feedback to a defensible roadmap. SignalDesk makes the evidence trail first-class: every opportunity, PRD line, and experiment links back to real user quotes.

ICP & MVP scope

ICP: early-stage SaaS / founder with 50–500 feedback items. MVP: import, cluster, ask-over-feedback, and one-click PRD + roadmap + experiment generation with citations. Out of scope: integrations with live ticketing tools and multi-user review workflows.

Metrics & guardrails

North star: validated opportunities converted into PRDs. The key guardrail, unsupported recommendation rate, directly protects trust, which is the whole value proposition of an evidence-first tool.

Resume bullets · AI Engineering
  • Built a RAG pipeline (ingest → chunk → embed → cluster → retrieve → generate) over user feedback with pgvector and source-grounded generation.
  • Enforced a citation guardrail (≥2 evidence links per recommendation) to eliminate unsupported AI claims.
Resume bullets · AI PM
  • Designed a research-to-roadmap workflow that converts raw feedback into cited PRDs, roadmap bets, and experiment plans.
  • Defined an evidence-first metric framework where the core guardrail (unsupported recommendation rate) protects user trust.