Ayush Kumar Ayush Kumar

Generic RAG falls apart on legal text, where one answer needs clauses stitched together from different documents. We built on LLMCompiler, so a planner lays out a DAG of retrieval tasks up front and independent sub-queries run in parallel.

Retrieval goes through Voyage-Law2 and Cohere reranking, then an auto-merger that rejoins clauses split mid-chunk. A joiner grades whether the evidence actually answers the question and replans if not, falling back to a beam-search retriever for multi-hop chains, then to a human. Claude and OpenAI back each other up so one outage can't take it down.

When an airline changes its schedule, thousands of passengers have to be moved onto other flights without breaking PNR priority, cabin class, connection times or special passenger needs. For Mphasis's problem at the 12th Inter-IIT Tech Meet, we built a rule engine that scores every PNR and every candidate flight, then solves the reassignment as a mixed-integer program.

We solved it classically with Gurobi and as CQM and BQM formulations on D-Wave's quantum annealer, and sped up preprocessing with prune-and-search, multithreading and dynamic programming. The rules live in a Streamlit UI, so an operations team can retune them without touching code. It took gold, first of 20 IITs.

Scanned medical PDFs mix lab tables, prescriptions, discharge notes and X-rays on the same pages, and one generic prompt kept hallucinating structure. I split the job across 14 narrow MedGemma agents on Vertex AI, plus a YOLOv8 model I fine-tuned on a 301-image dataset I annotated myself to find scans inside PDFs. Every detected crop gets a second opinion from an LLM before it's accepted.

A shared report summary feeds the downstream agents so nothing re-reads raw pages, and a layered JSON repair chain keeps structured output from breaking. In the UI you can draw a box on a scan and ask about just that region.