RECOMMENG1NE

Build your first engine
Research

The case for provenance-first AI.

An answer you can't defend isn't a decision — it's a guess with better grammar. This is the thesis Recommengine is built on: what goes wrong when AI recommendations arrive without provenance, what a traceable recommendation actually contains, and how we measure whether ours hold up.

Try the live demo → See the platform
The problem

Decisions are scattered, slow & hard to trust.

Context is fragmented
Signals live across chat, calendar, maps and files — never in one place at the moment you decide.
AI you can't trust
Black-box answers with no provenance are hard to act on — you can't see how the recommendation was reached.
Monitoring isn't deciding
Dashboards show everything and recommend nothing. The result is more data, not better decisions.
The proof

This is what a traceable recommendation looks like.

No black box. Every answer ships with its execution trace — which agents ran, what workflows and tools they used, which context sources they read, and where policy gates approved the calls. If you can't defend it, don't act on it.

See it with your own data →
What we measure

Traceability is only useful if something checks it.

Every recommendation passes a quality gate

Before a recommendation reaches you, it is evaluated against the goal and constraints the agent was given. Recommendations that can't show their work, or that drift outside their constraints, are flagged rather than shipped quietly.

Every tool call is policy-gated and audited

Agents don't call tools directly. Each invocation passes a policy gate that records who asked, what was approved, how long it took and what came back — so the trace isn't a summary written after the fact, it's the execution record itself.

Your decisions feed the next recommendation

Accepting, snoozing or dismissing a proposal is signal. Those decisions build a profile that reprioritizes what surfaces next, so the engine converges on what you actually act on instead of what merely looks urgent.

Isolation is a correctness property, not a setting

Every agent, engine, workflow run and context source is scoped to its owner. Cross-tenant context can't leak into a recommendation, because there is no code path where one owner's context reaches another's engine.

See a trace with your own data.

Compose an agent, ask it something that matters, and read the execution path it hands back. Start in minutes.

Create a free account → See the platform