Trajectory: A Loan Assessment You Can Audit

What it's trying to do

Trajectory began at UofTHacks with a simple question: what would a fairer first look at a small-business loan require? Lending decisions tend to lean on a narrow set of signals, and the businesses that sit outside them are often the ones that get overlooked. We wanted an assessment that considered more of the picture.

Just as important, we wanted the assessment to be explainable. A single prompt that returns "approve" or "decline" is quick to build and impossible to question, and that felt like the wrong shape for a decision that affects someone's business.

How it works

Instead of one general model call, Trajectory splits the assessment across specialist agents built with LangChain and Gemini. Each one focuses on a different kind of evidence: the business's finances from connected banking data through Plaid, the market it operates in, and its location using Google Maps context. Because those analyses don't depend on each other, they run in parallel.

Raw transactions and map results aren't useful to a model on their own, so each source is first shaped into a structured state before any reasoning happens. The final assessment is built from those intermediate results rather than generated from scratch, and the intermediate results are stored, so after a run you can go back and read what each specialist concluded.

What worked

• Keeping the reasoning. Storing each agent's intermediate state turns the result into something a reviewer can inspect, not just trust.
• Specialists in parallel. Splitting the work by evidence type kept each agent's job focused and kept the overall assessment quick.
• Structuring the inputs. Shaping raw financial and location data before prompting gives each agent a consistent, readable picture to reason over.
• Taking security and testing seriously for a hackathon project, with isolated handling of banking tokens and a 33-test suite that passes in full.

The limits worth naming

An auditable trail helps a person review an assessment; it doesn't replace that person. The agents reason with language models, and their conclusions are only as good as the data they receive and the way they are prompted. Trajectory is best understood as a tool that organizes evidence for a human decision, not one that makes the decision.

The test suite has similar boundaries. It covers the system path, meaning the data flow, storage, and assembly of the assessment, but tests can't tell you whether an agent's judgement about a business is actually sound. That kind of evaluation needs real cases and people who understand lending, which is the natural next step beyond a hackathon build.