Give the agent a concrete output contract.
I led the engineering while defining the shape of the research result. LangChain tool-calling loops combined investor research and signal extraction into structured evaluations.
4Engineers led
4 moWorking proof-of-concept cycle
THE PRODUCT TODAY
Visit Boxsy See it in the world.
This is the current public website. The case study below focuses on my contribution during the project.
Start with the
actual problem.
Investor research is useful only when someone can act on it. A fluent AI response was not enough: the output needed a reliable structure and a path into the databases and workflows that pilot customers already used.
The work behind
the outcome.
Treat retrieval and output handling as product work.
I built a Perplexity retrieval layer and self-correcting output handling around GPT-4 and Claude. Schema-valid JSON was the target contract, with parsing and correction built into the flow.
Do the last mile with the customer.
I worked with pilot VC teams to integrate private investor databases and built adapters for their tools. FastAPI endpoints included fallback and throttling behavior for unreliable third-party services.
How the pieces connect.
01Investor databases
02Retrieval + research tools
03Agent reasoning loop
04Validated evaluation
05Customer tools
Explanatory system view · reconstructed from the project
Hands on.
People with me.
I led four engineers and wrote code alongside them, owning both the product contract and the integration work needed to put the output into real customer workflows.
LangChainPerplexity retrievalFastAPISchema validationCustomer integrations
What came out of it.
The team delivered a working proof of concept over four months, with external VC pilot teams using the system against their own investor databases.
CONTEXT & LIMITS
Boxsy was a Purdue university project, not an employment role. Pilot usage is not a claim of enterprise-wide deployment, and structured output handling is not an absolute reliability guarantee.
Caterpillar
Paint on the line. Clarity in the data.