Rohith Sai Kanchi

About Rohith Sai Kanchi

I build ML pipelines and Generative AI systems that solve real enterprise problems — not demos, but production systems handling real data at scale.Over the past 4 years at Techfinite Systems, I\'ve worked across two distinct problem spaces. On the classical ML side, I built an end-to-end claims prediction pipeline for SS & C — training XGBoost and LightGBM classifiers to automatically route claim outcomes (approved, denied, escalate) across 10+ source formats, achieving an 88% F1-score on a heavily imbalanced dataset. That project also included a FAISS-backed RAG pipeline with GPT-4 so insurance adjusters could query 40-page claim documents in natural language instead of reading them manually.On the GenAI side, I\'ve been contributing to Yultra-AI — an enterprise multi-agent AI platform where domain-specific LLM agents handle Business Capability Modeling, Risk, and APM workflows. The work I\'m most proud of is the MCP (Model Context Protocol) SSE server integration — connecting LLM agents directly to live Java/Spring Boot backend APIs so responses are grounded in verified system data, not model assumptions. Combined with Qdrant-based hybrid retrieval and OAuth 2.0/OIDC-secured endpoints, it\'s a production-grade GenAI system, not a prototype.Before Techfinite, I built my foundations at Yorosis Technologies — Python pipelines, classification models, document parsing, CI/CD workflows.What I work with: Python · XGBoost · LightGBM · LangChain · OpenAI GPT-4 · RAG · Qdrant · FAISS · MCP · Prompt Engineering · Docling · Flask · FastAPI · Docker · Kubernetes (Azure AKS) · MongoDB · PostgreSQL · OAuth 2.0 · GitHub ActionsOpen to Data Science Engineer and ML/GenAI Engineer roles. Feel free to connect or message me directly.

Experience

1. Data Science Engineer Ii

   Techfinite Systems

   Aug 2022 — Present · Baltimore, MD, US

   Project 1: SS & C – Claims Prediction & Data Matching• Contributed to building data ingestion and chunking pipelines to normalize structured claim records and unstructured supporting documents (medical reports, EOBs) across 10+ source formats; wrote Python preprocessing scripts for field standardization and feature engineering ahead of downstream model input.• Implemented fuzzy matching and entity resolution logic using RapidFuzz (Jaro-Winkler, token-sort ratio) to reconcile provider names, policy IDs, and diagnostic codes across inconsistent source data; this helped reduce data mismatch errors and improved downstream triage accuracy noticeably.• Trained XGBoost and LightGBM classification models to predict claim outcomes (approved / denied / escalate) using engineered features from historical claims and ICD-10 code embeddings; managed data labeling workflows in Label Studio with iterative review cycles, achieving an 88% F1-score on the held-out test set.• Helped design prompt engineering templates for an LLM-based claims summarization feature (GPT-4 via Azure OpenAI) enabling adjusters to query claim histories in natural language; contributed to a FAISS-backed RAG pipeline that reduced the information lookup steps in the adjuster review process.• Assisted in containerizing the inference service with Docker and supporting its deployment on Azure AKS; helped configure MongoDB collections for storing prediction outputs and audit logs following team-defined schema patterns.