Inspiration
Ask three AI models the same question and you get three different answers, each delivered with the same confidence. Most people pick one and hope. We wanted the opposite: ask several independent models from different makers, then show plainly where they agree, where only one of them is talking, and which claims your own sources actually back.
What it does
KENNEL sends one question to three open models served on Nebius Token Factory, one per maker: NVIDIA Nemotron 3 Super, Alibaba Qwen3 30B and Google Gemma 3 27B. Each answers independently in a few plain sentences. KENNEL then:
- Reconciles the answers into claims, finds the ones at least two models share (agreement) and the ones only one model said (dissent), and shows an agreement meter.
- Grounds every claim against sources you paste in: each claim is tagged "source s1" or "no source", using a content-word coverage check that never lets a single shared word count as support.
- Learns from votes: mark an answer Helpful or Not helpful and that model's smoothed weight updates (Laplace-smoothed).
Typical round trip: about 2 seconds for all three models.
How we built it
- Python 3.12, standard library only, built from an empty folder during the hackathon window (first commit 2026-10-28T19:06:03Z).
- Runtime: AWS Lambda Function URL calling the Nebius Token Factory OpenAI-compatible API (
https://api.tokenfactory.nebius.com/v1), modelsnvidia/nemotron-3-super-120b-a12b,Qwen/Qwen3-30B-A3B-Instruct-2507,google/gemma-3-27b-it, in parallel threads. - A short system prompt keeps answers to 3 to 5 plain sentences (max 300 tokens, temperature 0.2), which took the first live run from 60 s down to about 2 s.
- Reconciliation strips markdown, protects abbreviations like "e.g." from sentence splitting, and groups claims by token overlap.
- 18 automated tests run with fake model calls, so no key or network is needed to test.
- The API key lives only in the function's environment, never in the repo. Concurrency is capped at 3 requests at a time.
Challenges we ran into
- The first live call took 60 seconds: models wrote long markdown essays and one timed out. Fixed with the plain-sentence system prompt and per-model token and time limits.
- Naive sentence splitting turned "e.g." into a claim called "g". Fixed with abbreviation protection and a minimum claim length.
- Jaccard similarity against a long source passage scored real support at 0.1. We switched to coverage of the claim's own content words, with a two-shared-word minimum.
- Two of the model IDs we first planned did not exist on the account; we checked the live model list and picked three that answered a live test.
Accomplishments that we're proud of
- A complete product, not a notebook: a public demo page with an agreement meter, source tags, per-model cards and voting.
- Three different makers in one crowd, with NVIDIA Nemotron at the centre.
- Every claim either backed by your source or clearly marked as not.
What we learned
Agreement between models only means something when the models are genuinely different and are checked against evidence. Fast, short, plain answers reconcile far better than long ones.
What's next for KENNEL
- Use vote weights to rank answers and weight the agreement score.
- Keep votes in permanent storage.
- More makers in the crowd, and a long-document mode that reads big sources in slices.
Feedback on Nebius Token Factory and NVIDIA models
The OpenAI-compatible endpoint meant zero client changes between models, and the live /models list made it easy to confirm what was actually served. Nemotron 3 Super gave careful, well-hedged answers; it was the slowest of the three (about 4 s on the first call versus under 1 s for the others), which mattered for an interactive crowd, so per-model timeouts are worth documenting prominently.
Licensed by NVIDIA Corporation under the NVIDIA Open Model License.
Log in or sign up for Devpost to join the conversation.