Guardrail Labs (RuleLock)

Inspiration

Most budgeting apps tell you that you overspent after the money is already gone.

RuleLock works before the decision.

People often set good financial rules when they’re calm, then break them when they’re rushed, tempted, or emotional. We built RuleLock to turn those rules into real-time guardrails.

What it does

A user can write:

“Keep my dining spending under $200 per month.”

NVIDIA Nemotron converts that sentence into a structured financial rule. Our deterministic TypeScript engine then checks purchases against that rule.

If the user has already spent $188 on dining and tries to spend $31.84 at Chipotle, RuleLock calculates $219.84 in projected spending, detects that it exceeds the $200 limit, and returns:

HIGH RISK
DELAY

For high-risk moments, ElevenLabs turns the warning into a spoken intervention. Users can cancel or override the purchase, and the decision is recorded in their scorecard.

How we built it

We used Next.js, TypeScript, Tailwind CSS, NVIDIA Nemotron, ElevenLabs, and Vercel.

Our key design decision was simple:

Nemotron understands what the user meant. Deterministic code decides whether the transaction violates the rule.

We never rely on the LLM for financial arithmetic.

Challenges we ran into

Our first Nemotron parser scored only 5/10 (50%). We found malformed JSON and ambiguity around phrases like “below,” “under,” and “less than.”

We added structured output, clearer operator rules, validation, and retry logic. On the same benchmark, the parser improved to 10/10, while our deterministic rule engine also scored 10/10.

Accomplishments that we're proud of

We built an end-to-end system where Nemotron interprets intent, deterministic code evaluates financial risk, ElevenLabs delivers high-risk voice interventions, and the user's decision flows into a live scorecard.

We’re especially proud that we tested the AI, found real failures, and improved it rather than assuming the model would always behave correctly.

What we learned

The biggest lesson was that LLMs are strongest when they have a narrow role inside a larger system.

Nemotron handles language. TypeScript handles financial logic. ElevenLabs handles the intervention.

What's next for Guardrail Labs (RuleLock)

A future version could connect to real bank or card data, support household guardrails, and surface longer-term behavioral insights.

The model doesn’t decide what I should spend. I do. RuleLock just remembers my decision when I’m most likely to forget it.

Built With

• api
• css
• elevenlabs
• finance
• financial
• generative
• github
• localstorage
• nemotron
• next.js
• nim
• node.js
• openai
• rest
• sdk
• tailwind
• technology
• text-to-speech
• typescript
• vercel