What jugeni is
cat READMEOpen-source agent infrastructure for the parts most AI products skip until something breaks in production. Not a chatbot. The bookkeeping underneath one: what it decided, what it must not forget, what it refused, and on what evidence.
Source-attributed atoms
Per-atom provenance, not flat conversation history. Every fact knows where it came from and when it goes stale.
Explicit lifecycle
Proposed → accepted → locked. Defended locks enforced at admission, not silently overwritten at retrieval.
Process continuity
Work threads that survive across sessions: state, next step, and the trail that got there.
Write-time gates
Reject confident-but-unverified output before it propagates. Refusal is a first-class result, not a failure.
agile4ai
the working frameFive primitives I keep coming back to across every engagement. The parts that read as overhead until the day they are the only reason you can answer for what your system did.
- Auditable decisions with an explicit lifecycle, not silent overwrites.
- Defended locks on what must not move, enforced at admission time.
- Source-attributed memory with per-atom provenance, not a flat log.
- Write-time invariants that reject confident-but-unverified output.
- Refusal as first-class output. "I will not answer this" is a feature.
More on it post by post → dev.to/jugeni
Receipts
public work, every number carries its queryWho runs this
./whoisMike Czerwinski. Twenty years building software: eleven writing code, nine running delivery across banks (BNP Paribas, Citi), automotive R&D, and travel-tech. Now building and auditing AI agent stacks: memory, decisions, locks, and the verification layer between a model and its consequences. Based in Poland, working in English and Polish.
Three product lanes, one obsession: making opaque AI legible enough to ship without lying to the user. Content provenance (sentixx.net), quantitative trading research (ziom), custom personal AI (jessi).