ANUSH
SOMASUNDARAM

Built a multi-agent CrewAI pipeline that automated literature review end-to-end — query expansion, retrieval, relevance screening, and summarization — replacing a fully manual process. Screened 46,000+ papers across ACM, IEEE, Springer, and other libraries down to a targeted ~200 for manual review, with a cross-model consensus mechanism (two LLM-backed agents voting on relevance) added as a validation layer at the screening gate.

Designed a proof-of-concept automatic number plate recognition (ANPR) system end-to-end — camera capture through OCR inference — using YOLO for plate detection and PaddleOCR for recognition. Built a confidence-based sanity-check module with regex format guardrails that lifted accuracy from ~70% into the high 90s without retraining. Wrote a PaddleOCR fine-tuning guide viewed 32,000+ times.

Contributed to UpTrain's open-source LLM evaluation tool, supporting A/B testing and prompt-based LLM-as-a-judge pipelines. Owned a Plotly-based visualization layer for side-by-side comparison of evaluation runs, shipped back into the open-source library. Documented internal codebases and ghostwrote technical content for the company blog.

Built an Industry Relation Mapping system on the National Industry Classification taxonomy, using spaCy NER to extract entities from articles and a custom Cypher query generator to build the graph without hand-written queries — capturing 15,000+ relationships across 2,300+ industries. Rebuilt the pipeline in 2025 as an agentic workflow using CrewAI and LangChain.