Your company’s work.
Your own intelligence.
In development · Direction published October 28, 2026
“This deployment is failing. Have we seen this before?”
- Deployment runbook
- Previous incident review
- Current release notes
- Establish the facts.Bring together the relevant changes and prior fixes, with source references.
- Make uncertainty visible.Separate evidence from hypotheses and identify the next checks.
- Keep the reviewed outcome.Record what actually worked and use it to test future improvements.
Your selected sources → your approved model → a brief your team can review.
Start with the next investigation.
A deployment fails. The answer may be spread across a runbook, a past incident, a code change, and someone’s memory. Our first workflow brings that evidence into a reviewable brief.
Potluck will identify current facts, related fixes, hypotheses, missing evidence, and suggested checks. Sources stay attached so a reviewer can inspect the basis for a recommendation. The first version focuses on investigation and handoff.
The outcome we want to measure is time to an accepted, useful brief, alongside citation accuracy and the corrections it needs. Customer results have not yet been established.
Build on knowledge you can trace.
Company Brain starts with selected source collections, explicit exclusions, document revisions, and clear distinctions between current procedures, historical decisions, and proposed changes.
Workspace permissions, source updates and deletion, audit history, and controlled access are part of the plan. Shared customer use depends on those controls. Existing personal memory is not an enterprise permission system.
Choose where the intelligence runs.
Potluck already runs open models locally and connects owned machines. We’re building on that foundation so organizations can choose approved models and execution locations, from one workstation to private infrastructure.
The mesh remains part of Potluck. Experimental model splitting explores additional capacity when a workload needs it. The first company workflow can start with one model on one approved machine.
Turn reviewed experience into tested improvements.
A reviewer can accept, correct, or reject a brief and record what actually happened. The planned next step is a proposed knowledge update and a reusable evaluation case, with provenance and review.
Improvement may come from better retrieval, procedures, prompts, routing, or model selection. Specialized models and adapters come later if approved examples and independent evaluations show a benefit. Corrections will not automatically trigger training.
Bring us a real workflow.
Investigating recurring failures? Reconstructing past decisions? Tell us how your team does that work today and whether you would help test this direction.