ForgeNine

AlphaCase study

Sparks: company memory that shows its sources

Sparks captures what experts know and answers with the evidence. It is in alpha, and it shows how I design AI that you can check.

By Rob Taroncher, ForgeNine

Join the course waitlistSee how I check answers

One expert, one calendar, one conversation at a time

Answers often live in one person's head.

Glossary

Customer churn

A customer stops renewing.

Sparks

Customer X left after the August renewal-policy change.

  • event
  • timing
  • decision
  • official source
  • corrections
Illustrative example, not customer data.

What I built

An AI teammate you teach on purpose.

  • Teach it like a colleague

  • Interviews that find gaps

  • Documents become knowledge

  • Answers show their evidence

  • Memory over time

  • Private, team or company

  • Linked to tasks and projects

  • Works with several AI models

How it works

From what people teach it to an answer you can check.

  1. 1

    Capture

    docs, interviews, corrections

  2. 2

    Digest

    summarize, link, index

  3. 3

    Remember

    sources, graph, timeline

  4. 4

    Answer and check

    checks evidence, cites sources or says it doesn't know.

Corrections feed the next answer.

Simplified view.

Under the hood

  • Next.js 16
  • TypeScript
  • Prisma, 78 data models
  • PostgreSQL with pgvector
  • Neo4j
  • Anthropic Claude
  • Databricks model serving
  • Google Vertex AI
  • LM Studio
  • llama.cpp
  • BAAI bge-m3
  • MLflow tracing
  • Jest

How I check its answers

Sounding right and being right are different tests.

  • Two scores per answer

    One for correctness, one for staying inside its sources.

    Overall010.905
    Grounding010.983

  • Hidden test questions

    60 cases. 12 sealed with SHA-256, hidden during tuning.

    • Visible
    • Sealed

  • Best full run so far (not yet good enough to ship)

    The best recorded run scored 0.905 overall and 0.983 on grounding, with 53 of 60 cases passing.

    • Passed

  • Public benchmark

    LongMemEval 10-question slice, April 10, 2026: 8 of 10 passed.

    0.90 Recall@5

Real screens, demonstration data

Captured May 2026. Details may have changed.

Sparks interview view during an active session: the assistant asks who the key people on your team are, with a sources panel and a captured insights panel. Demonstration data.
Interview session
Sparks interview history listing three active interview sessions, with an empty details pane. Demonstration data.
Interview history

What's next

Planned, not built yet.

  • Join meetings as disclosed AI
  • Speak only from what it learned
  • Hand judgment calls to the person

How it was built

I built it with AI coding agents and checked their work.

Read more about Rob

  1. First commit

  2. Personal-assistant fork, 11 commits

  3. Latest of 1,413 commits

What I learned

  • Quiet workarounds add up

  • Update docs with the code

  • Write plans down for AI

  • Check answers against sources

An AI answer is a draft until you see its source.

Learn to check AI answers like I do

We'll email you when the first lesson is ready.

Join the waitlistSee the nine skills