-
-
Logo
Runway — The Smoke Detector for Your Business
Inspiration
Small businesses rarely fail because their owners are not working hard enough. More often, problems develop quietly across places that are difficult to monitor at the same time: a customer pays an invoice late, a supplier raises prices, revenue starts trending downward, payroll is approaching, or an important customer suddenly announces layoffs.
Individually, none of these events necessarily looks catastrophic. Together, they can dramatically change a business's financial outlook.
The problem is that most financial tools tell owners what already happened. Dashboards show balances, accounting software records transactions, and spreadsheets help build forecasts, but the owner still has to connect the dots.
We wanted to build something different.
Runway is an AI-powered early-warning system for small businesses — essentially a smoke detector for their finances.
Instead of waiting for an owner to discover a cash-flow problem, Runway continuously turns financial activity, invoices, documents, and external business signals into actionable warnings. It answers three simple questions:
- What changed?
- Why does it matter?
- What should I look at next?
Our goal was not to build another financial dashboard or generic AI chatbot. We wanted to build a system that could detect weak signals early enough for a business owner to actually do something about them.
What We Built
Runway combines structured financial data with unstructured business information to maintain a continuously updated view of a company's financial runway.
Our demo follows Maya's Catering Co., a small business with:
- $43,200 in current cash
- $19,400 in expected inflows
- $51,700 in expected outflows
- $10,900 in projected ending cash
- an 18-day cash runway
Runway identifies risks that might otherwise live in completely different systems, including:
- an 18% supplier price increase
- a $12,400 invoice that is 17 days overdue
- an 11.8% decline in weekly revenue
- an upcoming $18,400 payroll obligation
- external news indicating that an important customer is experiencing layoffs
Each warning traces back to evidence so the owner can understand where it came from instead of simply trusting an unexplained AI prediction.
From Documents to Financial Impact
One of the most important workflows is live document analysis.
For example, Maya receives a notice from FreshFields Produce announcing a 7% surcharge, effective October 1, on her current $1,850 weekly spend.
The document is uploaded directly into Runway.
NVIDIA Nemotron extracts the relevant facts and supporting evidence from the unstructured notice. Runway then validates those facts and passes them to its deterministic financial engine.
The financial calculation itself is deliberately not performed by the LLM.
The estimated monthly impact is calculated as:
$$ \$1,850 \times 0.07 \times \frac{52}{12} = \$561.17 $$
Because the surcharge only applies to part of the active forecast period, Runway prorates the impact:
$$ \$561.17 \times \frac{18}{31} = \$325.84 $$
The forecast automatically changes from:
- 18 → 17 days of runway
- $51,700 → $52,025.84 in expected outflows
- $10,900 → $10,574.16 in projected ending cash
- $4,800 → $5,125.84 in projected reserve shortfall
Importantly, Maya's actual cash remains $43,200. Runway understands that a future supplier surcharge changes the forecast, not the money currently sitting in the bank.
Beyond the Chatbot
We designed Runway around a simple architectural principle:
AI understands. Code calculates. AI explains.
This became one of the most important decisions we made.
LLMs are excellent at interpreting messy information such as supplier notices, emails, and natural-language questions. They should not be the authoritative source for financial arithmetic.
Runway therefore uses NVIDIA Nemotron to interpret unstructured information and user intent while deterministic application code remains authoritative for balances, cash projections, scenario calculations, reserve shortfalls, and runway.
This creates a clear pipeline:
Source → AI extraction → Validation → Deterministic calculation → Financial state → Explanation
It also means every important warning can preserve provenance rather than becoming an unexplained AI-generated number.
Voice as a Financial Interface
Small-business owners are often working rather than sitting in front of a financial dashboard, so we also wanted Runway to work conversationally without becoming just another chatbot.
The Runway Voice Assistant lets an owner ask questions such as:
"Why did my runway go down?"
Browser speech recognition converts the question into text. Nemotron interprets the question against Runway's verified financial state and signals, and the response is validated against authoritative data before ElevenLabs turns it into natural speech.
The owner can therefore hear an explanation of what changed while still being able to inspect the underlying source and calculations.
Runway also supports multiple languages, including English, Spanish, French, Hindi, and Arabic, making the same financial intelligence more accessible to different business owners.
How We Built It
Runway uses a full-stack architecture designed around separation of concerns.
The frontend is built with Next.js, React, TypeScript, and Tailwind CSS. A FastAPI/Python backend owns the financial state, signal processing, document workflows, deterministic calculations, scenarios, and AI orchestration.
NVIDIA Nemotron, accessed through OpenRouter, handles unstructured signal extraction and grounded natural-language reasoning. ElevenLabs provides multilingual text-to-speech for the voice experience.
The production frontend is deployed on Vercel, while the API runs on Render.
We also built explicit contracts between the frontend and backend, validation around AI-generated structured output, provenance tracking, duplicate protection, deterministic fallbacks, and automated tests covering the application's financial and AI workflows.
Challenges We Faced
The hardest challenge was not connecting an LLM to financial data. It was deciding where AI should stop.
An early version of this type of system could easily ask an LLM to read a document, estimate its impact, and explain the result. That creates a dangerous failure mode: a convincing explanation can contain incorrect financial arithmetic.
We instead separated interpretation from calculation. Nemotron extracts facts such as a supplier, percentage increase, effective date, and current spend. Those facts must pass validation before deterministic code calculates their financial effect.
Another challenge was preventing double counting. A signal may already be represented in the baseline forecast, so discovering it again cannot simply add the expense a second time. We had to track which signals had already affected financial state and distinguish new information from existing assumptions.
Document extraction introduced another challenge: real LLM output is not always perfectly formatted. We built stricter prompts, structured validation, evidence requirements, confidence handling, and deterministic fallbacks so malformed AI output could not silently become authoritative financial state.
Voice introduced a different set of problems. We had to connect microphone recognition, grounded Q&A, verified financial context, multilingual responses, and ElevenLabs speech while ensuring that spoken answers could not invent financial values.
Finally, moving from localhost to production exposed the kinds of integration issues that only appear in a deployed system: monorepo builds, shared TypeScript contracts, environment configuration, cross-origin requests between Vercel and Render, and external AI-provider behavior.
What We Learned
The biggest lesson from Runway was that AI becomes more useful when it is given less authority in the places where correctness matters most.
Nemotron is extremely valuable for understanding information that traditional software struggles with: documents, language, intent, and relationships between signals. Deterministic software is much better at arithmetic and maintaining financial invariants.
Combining the two gave us something stronger than either approach alone.
We also learned that explainability is not just showing users an AI-generated explanation. Real explainability means allowing someone to trace a conclusion all the way back:
What source produced this signal? What fact was extracted? What calculation was performed? What changed in my forecast?
That principle shaped nearly every part of Runway.
What's Next
Runway currently demonstrates the system using synthetic financial and business data, but the architecture could eventually connect to accounting platforms, banking feeds, invoices, supplier communications, and other business systems.
Longer term, we envision Runway operating quietly in the background and alerting owners when combinations of weak signals begin creating meaningful financial risk.
The goal is simple:
Give small-business owners time.
Time to follow up on an overdue invoice.
Time to renegotiate with a supplier.
Time to adjust spending before payroll becomes a problem.
Time to act before a warning becomes an emergency.
Traditional financial software tells a business owner where they have been.
Runway is designed to tell them what's coming.
Built With
- claude
- codex
- deepseek
- hermes
- nemotron
- render
- vercel
Log in or sign up for Devpost to join the conversation.