QuantumCopilot
AI Loan & Financial Assistant
Ask about a loan in plain English and get calculated scenarios side by side, while the LLM never does the math.
- 6Evidence 1 · 6 loan productsLoan products the assistant can extract parameters for.Go to Resultsloan products
- Up to 3Evidence 2 · Up to 3 scenarios in parallelScenarios built from one question and calculated at the same time.Go to Key decisionsscenarios in parallel
- ~3xEvidence 3 · ~3x faster scenario stageAn estimate: three scenario calculations run in parallel instead of one after another, so the wait is the slowest call, not the sum.Go to Key decisionsfaster scenario stage
What's my EMI on a ₹20 lakh home loan over 15 years? And what if the rate drops?
Here are three rate scenarios for a ₹20,00,000 home loan over 15 years:
| Rate | Monthly EMI | Total interest |
|---|---|---|
| 9.0% | ₹20,285 | ₹16,51,360 |
| 8.5% | ₹19,695 | ₹15,45,062 |
| 8.0% | ₹19,113 | ₹14,40,348 |
Dropping from 9.0% to 8.0% saves you ₹2,11,012 in interest.
Make it 20 yearsCompare a car loan
The problem
People think about loans in sentences: “What’s my EMI on a 20-lakh home loan over 15 years, and what if the rate drops?” Loan calculators make them fill in forms instead.
Constraints
- The numbers must be exact. An LLM’s arithmetic isn’t good enough for money.
- Follow-up questions should feel natural.
- It’s a production API, so it has to be secure and keep working when dependencies fail.
How it works
- Question
- API key checkconstant-time compare
- Load conversation memorysql server
- LLM extracts typed parametersgpt-4.1-nano
- Build up to 3 scenarios
- Calculation engine, in parallelasyncio
- Comparison table + memory update
After an API key check and loading recent conversation memory, the LLM turns the question into typed parameters: loan product, amount, tenure, rate. Up to three scenarios are built and sent to the calculation engine at the same time, then the LLM writes a comparison table and the memory is updated. If no loan details are found, it falls back to a general finance chat.
Key decisions
- The LLM extracts, the engine computes. The model only turns language into typed parameters. A deterministic calculation engine does every number.
- Parallel scenarios. Up to three scenarios run at once, so the wait is the slowest one instead of all three added together. That made the scenario stage about 3x faster.
- Memory as summaries. Each turn is saved as a short summary. “Make it 20 years” changes only the tenure and carries everything else over.
- Boring security, done properly. Constant-time API key checks with key rotation, parameterized SQL, retries with backoff on rate limits and server errors, and one consistent error format.
Results
- Plain-English questions answered across 6 loan products.
- ~3x faster scenario stage, from running the scenario calculations in parallel.
- Live in production, running on 8 Uvicorn workers.
What I’d do next
- Bring the amortization engine fully in-house and benchmark latency formally.
- Move parameter extraction to schema-enforced structured outputs.