Inspiration

As a personal trainer I know the best that a wellbeing assistant is only useful if its advice can be traced. Most answers either come from a model's memory, with no source, or from a generic list. I wanted a council of experts where every point comes from a named source, and where any claim that can be checked on the web is checked before the user sees it.

What it does

The core is an expert council, run by a brain model hosted on Nebius Token Factory.

  • The brain (relay) reads the question and decides which experts to consult: diet, training, breath and stress, or sleep and lifestyle. It can consult several experts, follow up between them, and then write one answer.
  • Each expert answers only from its own knowledge bank: NHS, MedlinePlus, NCCIH and NIMH content, retrieved by embedding search (Qwen3-Embedding-8B on Nebius). An expert point is kept only if it cites a source the expert was given.
  • Tavily verification runs in two places:
    1. Web checks for experts. When a bank has too little on a question, the expert asks for a web search, and the results are added as cited sources.
    2. Food macros and calories. When a user logs a meal and states its calories, Tavily checks the figure against a web source. A mismatch is never changed silently: the user sees both numbers and chooses one.
  • Sources are shown under every answer. The answer says when no expert was consulted.

Supporting parts: a typed chat (voice is switched off in this build), food and activity logs, and a breathwork guide. They show the council in a daily-use setting.

How we built it

  • Brain and experts on Nebius Token Factory: Nemotron 3 Super 120B for the council, Nemotron 3 Ultra 550B for chat replies (plain prose), and Qwen3-Embedding-8B for knowledge retrieval.
  • Tavily for the web checks and the food verification.
  • Server: a Cloudflare Worker, with Vectorize holding the knowledge index and Durable Objects for per-device quotas.
  • Safety in code: crisis phrases get a fixed referral before any model call. Points without a source are dropped. The model marks health advice that could cause harm, and only those answers get the clinician line. ## Challenges we ran into
  • The first food check compared the wrong number: a per-cup figure was scaled to 200 g. The check now uses only figures stated per 100 g or per 100 ml, and it compares the median across sources. A single wrong site can no longer decide it.
  • The chat model wrote its planning into answers. Its reasoning is now off on the answer path, and a rule in the base prompt keeps replies to plain prose.
  • Medical questions are not grounded in a consulted expert in this build. The medical experts are off, so those answers say so instead of citing sources they do not have. ## Accomplishments that we're proud of That after trying to figure out how to build something meaningful, the hackathon pushed me to develop this ultimate solution to my problems. ## What we learned A council of narrow experts, each with its own sources, is easier to check than one large model answering alone. Keeping each model's job narrow (extract, cite, answer, verify) made every part testable, and the verification step caught real errors in our own food checks. ## What's next for Illunis Verify - Nemotron Nutrition Truth Nebius Token Factory, Nemotron, Qwen3-Embedding, Tavily, Cloudflare Workers, Durable Objects, Cloudflare Vectorize, Flutter, Dart, JavaScript (Node.js), GLSL

Built With

  • cloudflarevectorize
  • cloudflareworkers
  • dart
  • durableobjects
  • flutter
  • glsl
  • nebiustokenfactory
  • nemotron
  • node.js
  • qwen3-embedding
  • tavily
Share this project:

Updates

Submission history