Every month, a lender receives a loan tape: thousands of rows, one per loan per month. Nobody can read all of it, so teams check a small sample and hope the risky loans are in it. For the First Commit hackathon, my teammate Abhishek and I built LoanLens, which scores every loan the moment the file lands. It runs entirely on serverless AWS.
Architecture
Uploading a file is the only trigger.
- The browser gets a presigned URL and uploads straight to S3, so API Gateway's payload limit never applies.
- The S3 event goes through EventBridge to a Step Functions workflow.
- The workflow calls a container Lambda (arm64) running LightGBM models and a rule engine, and writes every scored loan to DynamoDB.
- If a run finds high-risk loans, Step Functions publishes to SNS using its native SDK integration, with no extra Lambda.
- The React dashboard on Amplify talks to API Gateway, which fronts Lambdas for the portfolio, a stress test, a policy gate and a copilot.
Nothing runs when nobody is using it: no EC2, no SageMaker endpoint, no OpenSearch.
Cedar for approvals
Who may approve a loan is a policy, not a pile of if-statements. I used Cedar, AWS's open-source policy language, running in Lambda through @cedar-policy/cedar-wasm . A junior underwriter is denied above their limit, and only the risk committee can override a flagged record.
Two things I learned: Cedar has no floating-point numbers, so risk and loan-to-value travel as integer percentages. And the cedar-policy package on PyPI is an unrelated placeholder; the official engine is the npm package.
A copilot that never hides which model answered
The copilot builds its facts first, from aggregates over every scored loan, and only then asks a model to explain them. It tries Bedrock first, then an OpenAI-compatible model whose key sits in SSM Parameter Store as a SecureString, then a deterministic template. Every answer says which one produced it.
Our new account stayed in Bedrock's account verification for the whole event. Building the copilot as a chain of providers meant the product kept working.
What went wrong
The most useful lessons came from failures:
- The Python 3.11 Lambda image is Amazon Linux 2, so current scipy and LightGBM wheels wouldn't install. Switching to the Python 3.12 image (Amazon Linux 2023) fixed it.
- DynamoDB rejects Python floats and NaN. Our first live run failed on exactly that, and I kept it as a regression test.
- Step Functions retries only Lambda service errors. A malformed upload is deterministic, so it fails once with a message naming the missing columns.
Cost
On-demand DynamoDB, 7-day log retention, a 7-day expiry on uploads and API throttling kept spend near zero on a $100 credit budget.
What I would do next
Add authentication with Cognito and load-test larger tapes. LoanLens is a prototype on synthetic data, and its output is a recommendation, not a decision.
Loading article