At PreduskTechnology, I engineered a production IAA framework in Python that automated consensus scoring across 5,000+ annotations, cutting QA effort by ~60%. I also built a speaker diarization pipeline on Sarvam Saras v3/v4, reducing DER to ~13%.
At NeoLen, I designed a LangGraph multi-agent workflow for real-time clinical decision support and built streaming voice agents that cut per-turn latency and improved response efficiency by ~35%. I also developed LLM-based extraction pipelines for clinical parameters in unstructured EHR notes, reducing consultation turnaround by ~30%.
In my projects, I built SatyaNetra, a multimodal deepfake detector that achieved ~92% accuracy, and DocMind, a RAG pipeline with a ~0.84 faithfulness baseline. I also created a CLI-based LangGraph agent with user- and project-scoped isolation, checkpointing, and cross-session recall.
