Northbridge Brief

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 Northbridge Brief

Inspiration

Avery and Morgan do not want another noisy finance chatbot. Northbridge is their fictional home—a place where bills, remittances, and admin anomalies pile up, and where notification fatigue is as bad as a missed duplicate charge.

Most AI money apps narrate every transaction. Northbridge Brief was inspired by a simpler contract: stay quiet by default. Investigate what is due, research what changed when something looks wrong, reason into one speakable brief, and escalate only when policy says HUMAN_REQUIRED. Routine fiber, rent, and in-band utilities should never interrupt dinner.

The Nebius × NVIDIA hackathon made that product real: Token Factory gives an OpenAI-compatible path to NVIDIA Nemotron, so we could split “fast plan” and “careful brief” across Nano and Super, stretch credits, and still ship a full Control Plane judges can click through offline.

What it does

Northbridge Brief is an agentic research/ops assistant for household admin:

1. Sees what is due — fixture household bills and funding for Avery & Morgan Quinn.
2. Detects anomalies — amount jumps, duplicate payee/amount/due, sensitive categories (medical/tax/legal), and funding below the reserve floor.
3. Researches — Tavily (or offline fixtures) for every escalated anomaly.
4. Reasons — Nemotron Super composes a plain-spoken, speakable decision brief.
5. Auto-handles safe bills silently (AUTO).
6. Escalates once when a real human decision is required (HUMAN_REQUIRED).

The live Control Plane shows the agent timeline (plan → tools → policy → Tavily → brief), ledger, and speakable briefs—not a thin chat wrapper.

Judge path (~60s, no keys):

git clone https://github.com/dl88jing/northbridge-nebius-brief.git
cd northbridge-nebius-brief
./demo.sh

Interactive: ./demo.sh dev → http://127.0.0.1:3000 → Run quiet pass.

Decisions: AUTO for Cascadia Fiber and Rent; HUMAN_REQUIRED for StreamFlix (price jump + research), both PG&E notices (duplicate pair), and Bayview Dental (medical + research).

How we built it

Stack

• Next.js App Router UI + Vercel serverless API routes
• Investigation agent (lib/agent.ts) with deterministic household tools
• Quiet policy gate (lib/policy.ts) — AUTO vs HUMAN_REQUIRED (not LLM-decided)
• Nebius Token Factory + NVIDIA Nemotron for plan + brief
• Tavily for live web research (fixtures when no key)

Model and research flow

The dashboard calls an investigation agent. Nemotron Nano creates the investigation plan through Token Factory. Deterministic tools inspect bills, funding, and policy. Tavily (or fixtures) researches anomalies. Nemotron Super creates the decision brief through the same OpenAI-compatible chat completions API.

Default models are nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B for fast planning and nvidia/nemotron-3-super-120b-a12b for reasoned briefs, via https://api.tokenfactory.nebius.com/v1/. Set NEBIUS_API_KEY for live Token Factory calls; without it, offline demo mode runs the same agent loop with Nemotron-shaped brief synthesis. Live failures fall back to offline briefs, so the product remains judgeable.

Policy gate

1. Amount outside expected band → HUMAN_REQUIRED + research
2. Duplicate payee/amount/due → HUMAN_REQUIRED + research
3. Sensitive category (medical/tax/legal) → HUMAN_REQUIRED + research
4. Funding below reserve floor → HUMAN_REQUIRED
5. Otherwise → AUTO (silent)

Best Use of Tavily

Every escalated anomaly is wired through the research tool. Live Tavily answers “what changed?” (for example, StreamFlix repricing, a PG&E duplicate FAQ, or dental remittance norms). Offline fixtures mirror the same queries so the Tavily path is always demonstrable. Research is not a sidebar—it feeds the Nemotron brief that Avery or Morgan would hear.

Challenges

• Credit-aware dual-model routing — Nano for planning and Super for briefs without making the agent loop feel inconsistent offline vs live.
• Deterministic policy vs LLM judgment — keeping AUTO/HUMAN_REQUIRED out of the model so quiet behavior is trustworthy and judgeable.
• Research that earns its place — Tavily must change the brief, not decorate it; we only research on escalate paths.
• Judge-friendly demos without keys — Token Factory and Tavily need credentials, so offline fixtures and Nemotron-shaped synthesis were mandatory for Devpost fairness.

Accomplishments

• A complete Control Plane product (timeline, ledger, speakable briefs), not a notebook POC.
• Escalate-only household ops with clear AUTO vs HUMAN_REQUIRED outcomes on first run.
• Dual Nemotron roles on Nebius Token Factory with a one-environment-variable live switch.
• Tavily (or fixtures) on every escalation, ready for the Best Use of Tavily bonus.
• One-command ./demo.sh judge path and a live Vercel deployment.

What's next

• Deeper Token Factory “agent profile” presets (Nano plan / Super brief as a first-class console pattern).
• Optional Nebius Serverless Jobs for asynchronous batch investigations.
• Richer multi-bill research trails and human-ack state in the Control Plane.
• Voice handoff of speakable briefs into the broader Northbridge home-ops surface Avery & Morgan already use.

Built With

• nebius-token-factory
• nvidia-nemotron
• python
• react
• tavily
• vercel