FlowFit — Verified Execution Infrastructure for AI Agents

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for FlowFit — Verified Execution Infrastructure for AI Agents

Inspiration

Business AI often stops at generating an answer. Real operations need more: evidence, policy checks, human approval for sensitive actions, verified execution, audit history, and recovery when something fails.

FlowFit is built around that gap between AI reasoning and dependable execution.

What it does

FlowFit turns messy business requests into controlled operational workflows.

A request enters as an intake. NVIDIA Nemotron on Nebius Token Factory converts the unstructured message into a normalized task. FlowFit then passes that result through separate operational layers:

Intake
→ Task Understanding
→ Context Matching
→ Decision Engine
→ AutonomyGate
→ Human Approval when required
→ ExecutionAuthority
→ Connector / Device execution
→ Recovery
→ Timeline / Audit

The key boundary is:

Model recommendation ≠ Action authorization

The model can interpret and propose. Deterministic policy decides whether work is blocked, requires review, or may proceed.

How we built it

For this hackathon build, FlowFit adds Nebius Token Factory as a first-class AI provider and uses:

nvidia/nemotron-3-super-120b-a12b

Nemotron currently powers structured task understanding and is also wired into the shared agent-model provider layer.

We added real runtime telemetry for provider, exact model ID, route, measured latency, and last successful call. The UI shows LIVE only after a real successful Nebius inference.

The public judge demo performs a real Nebius/Nemotron call and then applies a deterministic human-review gate.

Significant update during the hackathon period

FlowFit existed before this hackathon; the original Runner repository began in July 2026.

After the submission period opened on August 26, 2026, the project was significantly updated with:

1. Nebius Token Factory provider support.
2. NVIDIA Nemotron 3 Super task understanding.
3. OpenAI-compatible tool-call translation for the Nebius provider path.
4. Real runtime telemetry and a LIVE evidence panel.
5. Integration of the new runtime with FlowFit's existing Decision Engine, AutonomyGate, approval, ExecutionAuthority, and recovery architecture.
6. A public judge demo that runs the real Nebius/Nemotron inference path.

Challenges

The main challenge was keeping probabilistic model reasoning separate from operational authority. Instead of asking the model to be more careful, FlowFit uses explicit boundaries for decisions, approvals, execution, verification, and recovery.

A second challenge was making the hackathon integration verifiable. The runtime evidence panel is intentionally driven by actual successful calls rather than a static badge.

What we learned

Reliable business agents need different mechanisms for different jobs:

• LLM: interpret ambiguous human input.
• Context: connect the task to company data.
• Decision Engine: apply explicit rules.
• AutonomyGate: decide whether automation is allowed.
• Approval: preserve human authority where required.
• ExecutionAuthority: constrain real actions.
• Recovery: handle external failure instead of assuming success.

Nebius Token Factory made it straightforward to add an open NVIDIA model to the existing provider architecture without moving those control boundaries into the model.

What's next

FlowFit is a working prototype under active development, not a finished commercial product.

Next work focuses on deeper real-world connectors, reusable verified procedures, broader failure recovery, industry-specific modules, and longer-running reliability and security testing.

Built With

• ai-agents
• human-in-the-loop
• nebius
• next.js
• node.js
• nvidia-nemotron
• typescript
• workflow-automation