RIMAZ Nemotron Copilot

AI copilot for short-video creators: ad scripts in Darija, French and English, translation, hashtags and a weekly plan from a channel own stats - NVIDIA Nemotron on Nebius, budget-capped.

Comment

Inspiration

Short-video creators and sellers on RIMAZ, a reels platform, spend hours on work that is not creating: writing ad scripts, translating captions, picking hashtags, and planning what to post next. We wanted an AI copilot that does this busywork in the creator's own languages, so cheap to run that anyone can host it.

What it does

One page, four tasks:

  • Ad script: a ready-to-record ad script in Darija, French or English.
  • Translate: captions and text between these languages.
  • Hashtags: relevant tags for a post.
  • Weekly plan: a 7-day content plan built from the creator's own public channel stats on RIMAZ - and it says plainly when the data is too thin.

Every answer shows the exact Nemotron model id that produced it. With no API key configured, the app runs in template mode where each answer is labelled TEMPLATE - no model called: it never pretends a template is model output.

How we built it

  • Front end + API: a single Cloudflare Worker (JavaScript, no framework) serving the web page and four JSON routes.
  • Models: the NVIDIA Nemotron family served by the Nebius Token Factory chat completions API, routed by job: nemotron-3-super-120b-a12b for ad scripts, Nemotron-3-Nano-30B-A3B for translate and hashtags, Nemotron-3-Ultra-550b-a55b for the weekly strategy.
  • Budget authority: a Cloudflare Durable Object serializes every model call. Before each call it reserves the worst case (max_tokens + the real prompt's UTF-8 byte count + a 64-token framing allowance - byte count, unlike character count, upper-bounds tokens for byte-level tokenizers even on Arabic input), settles to actual usage after, and rejects oversized reservations instead of clamping them. A reservation is released only when the provider provably never billed the call. Hard caps: 40 calls/day, 5 strategy calls/day, 500k total tokens - a worst case of about $0.50, inside Nebius's $1 trial. Calls fail closed with 503 if the binding is missing.
  • Data: the strategy task reads the creator's public posts from RIMAZ's public API.

Challenges we ran into

  • Keeping a $1 trial safe: the trial does not stop spending (Nebius auto-charges the card past zero), so the caps had to be the real protection - serialized, worst-case first, fail closed.
  • Token accounting for Arabic text: character counts under-count tokens; UTF-8 byte counts bound them.
  • Honesty by construction: template mode lets anyone run the app without credentials, explicitly labelled as not Nemotron output.

Accomplishments that we're proud of

  • Live Nemotron output verified on all four routes, with the model id shown under each answer.
  • 19 offline tests (Token Factory mocked) covering routing, request shape, template mode on all routes, health mode, the page, and origin rejection.
  • Zero standing cost: the whole app runs on the Cloudflare Workers free tier.

What we learned

Practical model routing across three Nemotron tiers, and how to build a serialized budget authority that makes a funded trial genuinely safe to demo.

What's next for RIMAZ Nemotron Copilot

Richer strategy inputs (retention and posting-time signals), a voice-first flow, and per-creator budgets as RIMAZ opens to more channels.

Built With

  • cloudflare-durable-objects
  • cloudflare-workers
  • javascript
  • nebius-token-factory
  • nvidia-nemotron
  • rimaz
Share this project:

Updates

Submission history