1st runner-up at WebOsmotic: round one banned AI, round two required it

WebOsmotic, a software company in Surat, ran an on-site AI/ML hackathon in August 2025. It began with resume screening, and only 20 people were invited in. I walked out as 1st runner-up, which is a dignified way of saying second, and I will take it.

• called on-site

  20

• overall

  2nd

• endpoints in 2 hours

  3

• AI tools in round 1

  0

The format was delightfully contrarian. Round one banned AI entirely. Round two made it mandatory. Somewhere, a philosophy department is writing a paper about it.

We had two hours to build three API endpoints, with no AI assistance of any kind. Just you, the docs, and your muscle memory:

1. 1Ingest: accept documents and split them into chunks.
2. 2Retrieve: answer a query using only the uploaded documents.
3. 3Generate: pass the query and the retrieved context to an LLM and return its answer.

That is the skeleton of every retrieval-augmented generation (RAG) system, and building it without autocomplete reveals very quickly whether you understand each bone. The decisions that matter are the unglamorous ones: how large a chunk should be and how much neighbouring chunks overlap, how to rank chunks against a query, how many to hand the model, and what the endpoint says when nothing relevant exists. ("I don't know" is a feature, not a failure.)

Then the rules flipped. We were handed a broken production codebase and asked to bring it back to life, using an AI agent as our guide.

This round tests something quite different from writing code: reading an unfamiliar system quickly, forming a hypothesis about what is broken, and interrogating the agent's suggestions rather than pasting them in on faith. An agent can find the right file in seconds. It can also fix a symptom with tremendous confidence while the actual cause sits two files over, smirking. Telling those apart is the entire skill.

• Fundamentals first, tools second. Round one rewarded knowing how RAG works with no help. Round two rewarded using AI well on top of that. You need both, in that order.
• Small, testable pieces. Three separate endpoints are easier to build, debug and demo than one heroic endpoint that does everything.
• Review the agent like a teammate's pull request. Read it, run it, and check that it fixes the cause rather than silencing the symptom.