Inspiration
AI research tools hallucinate citations, invent numbers, and ship vibes. Trace is a deep-research copilot built on the opposite principle: every claim cites a source, every number is recomputed in a sandbox, and nothing ships until a human approves it.
What it does
Ask Trace a hard question and it runs a six-stage pipeline — PLAN, SEARCH, ANALYZE, DRAFT, VERIFY, human approval — and produces an audited research brief with a JSONL audit trail. Web search runs through a real Tavily client: a live search + extract call was verified Oct 25, 2026 through Tavily's free Researcher plan (direct Tavily account; the Nebius Builders credit is still pending). The client previously sent the wrong auth header and was fixed to Bearer auth the same night. Data-analysis code is safety-gated (allowlisted imports, no dangerous builtins, bounded timeouts) and executed in isolated Token Factory Sandbox microVMs. A verifier model re-checks every claim against its cited sources with confidence scores. The final brief only publishes after explicit human approval.
How we built it — NVIDIA Nemotron on Nebius Token Factory
All inference runs through Nebius Token Factory's OpenAI-compatible endpoint, using NVIDIA's open Nemotron 3 models with deliberate cost/quality tiering: PLAN + VERIFY use nvidia/nemotron-3-super-120b-a12b, DRAFT uses nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B, with an optional nvidia/Nemotron-3-Ultra-550b-a55b tier. A live inference call was verified against Token Factory on 2026-09-15.
Challenges
Token Factory model IDs are case-sensitive; Nemotron thinking tokens consume token budget; scoped permissions enforced per stage with approval gates before publish.
Accomplishments
Working 6-stage pipeline with real stage-to-model tiering, hash-chained audit trail, and 11/11 passing tests; live Nebius inference verified end-to-end; live demo site + 2:20 narrated demo video.
What's next
Full-pipeline live evals on Super 120B once credits land; expanding Tavily search + extract coverage to more source types; Token Factory Sandbox execution against the account's endpoint.
