Companies are automating more work with AI. The hard part is knowing where human review still improves the result.
Keep review that no longer helps and you lose productivity.Remove review that still catches important mistakes and you create risk.
Janus learns where that boundary is over time and adapts the workflow.
The answer keeps changing
People adapt to AI. Models change.
A review that once caught important mistakes can become routine approval. A review that looks redundant can still catch rare failures.
AI also changes what professionals keep doing themselves: what they practice, what they delegate, and where they come to rely on the model.
So the handoff between person and model is not something you set once.
It has to be learned again as the relationship changes.
The same approval can mean four different things
A person agrees with the AI:
Normal decision log
01
they had already reached the same conclusion;
02
the AI corrected a mistake;
03
the AI pulled them away from a correct answer;
04
they never formed an independent view and simply approved.
Human agreed with AI
In a normal decision log, all four can look the same: human agreed with AI.
A single event rarely tells you enough about the relationship behind it.
Janus learns from the longer history: how this person reasons with AI, where reliance grows, what kinds of errors they still catch, how past cases turned out, and what changes when the interaction changes.
When independent judgment matters, Janus can change the next case too — for example by showing evidence before the conclusion, asking for an alternative, or asking the person to form a view first.
The point is not to add friction everywhere. It is to use independence, challenge, and automation where they add value.
What Janus does
Janus sits inside AI-assisted work as a persistent professional counterpart.
It builds and revises a private model of how that professional works with AI over time: evidence thresholds, confidence and calibration, reliance patterns, recurring blind spots, response to challenge, and where human judgment still adds something the model does not.
That model is made of contextual, revisable hypotheses — not permanent labels. Later cases can strengthen, weaken, or replace them.
Persistent counterpart
- Past case
- Past case
- Later case
Contextual, revisable hypotheses
- MORE HELP
- HOLD SOME HELP BACK
- CHALLENGE A CONCLUSION
- DELIBERATE PRACTICE
- RECOMMEND REMOVING THE HUMAN STEP
Janus uses the current model to adapt the next interaction. It can give more help, hold some help back, challenge a conclusion, create deliberate practice, or recommend removing the human step where review no longer earns its cost.
The same model on the same kind of case may call for a different handoff for a different person — or for the same person later.
What already runs
Longitudinal inference system
The longitudinal inference system behind Janus is already deployed.
It haspersistent state across sessions,evidence-backed versioned hypotheses,prediction and revision,provenance,and inspectable history.
It was built for an earlier Janus product centered on long-term AI personalization.
The enterprise version narrows that machinery to professional Human+AI work.
Where Janus fits
Janus needs repeated decisionswhere both a person and AI affect the outcome,the interaction can change,getting the handoff wrong matters,and reality eventually tells you what happened.
The economic fit is strongest when expert attention is scarce.
Security incident investigation is the first concrete workflow being built around those conditions.
Built by Egor Chirkunov
For eleven years I ran my own production company; the clients included Xerox, Microsoft, Beeline and Rostelecom. I built and led a narrative organisation of around fifty people, and since then I have built research systems, competitive observation from open data, and tools that turn difficult manual work into a repeatable process.
Janus needs those at once: reading how people reason and decide, holding plausible explanations against their alternatives without hiding where the evidence stops, and turning what that produces into something that runs. Since early 2026 I have been building it hands-on with AI-assisted development tools.
Janus is a one-person company today.