Back to projects
Jun 2026 – Present · Live · In progress
Janora — AI Agent Readiness Platform
Creator and sole developer
AI/MLWeb
The problem
Teams are shipping AI agents into real workflows faster than they can answer a basic question about them: is this one safe to turn on? The honest answer usually requires knowing three separate things — whether the agent has its facts right about the company, whether it behaves the way it claims to, and whether the actions it can take are safe to let it take. Janora exists to turn that from a judgement call into a decision with evidence behind it.
What I built
- Built a working v1 that tests whether an AI agent is safe to launch for a specific workflow, evaluating company truth, agent behaviour, and action safety.
- Designed workflow packs and agent manifests as the two inputs the system reasons over — one describing what a workflow actually requires, the other describing what an agent claims it can do.
- Built generated test scenarios on top of those inputs, so coverage comes from the workflow definition rather than from someone remembering to write the cases.
- Implemented three testing modes — manual, transcript, and sandbox — so an agent can be evaluated before it has ever run, from logs of it running, or in a contained environment.
- Built the certification status system and the Launch Readiness Report, which is the piece that turns a pile of test results into a clear go/no-go.
What I had to learn
- Designing an evaluation framework where no standard one exists — deciding what "safe to launch" even decomposes into was most of the work, and it had to be settled before any code was useful.
- Product architecture end to end as one person: schema, evaluation logic, and the report layer all have to agree on the same model of a workflow.
- Turning a fuzzy question into a decidable one. A report that says "mostly fine" is worthless to someone who has to decide whether to ship.
Results
- Working v1 live at janora.dev.
Stack
PythonTypeScriptPostgreSQL