Yathish Sadashivareddy

Built and published ConfidentLM, a production-grade Python library for LLM guardrails and uncertainty quantification using log probability analysis and confidence calibration (Isotonic and Platt scaling) with LangSmith-compatible tracing hooks, improving AI agent reliability by 27% on GAIA benchmarks. Designed and deployed agentic orchestration pipelines using LangGraph with ConfidentLM integrated into Smolagents and PydanticAI frameworks, implementing self-regulating decision loops based on uncertainty thresholds, raising task accuracy from 55% to 70% while enabling debuggable, auditable reasoning traces. Engineered a Human-in-the-Loop automation system for high-volume text classification, routing uncertain predictions to human reviewers boosting accuracy from 62% to 86% and cutting manual workload by 35%. Shipped ConfidentLM as a production microservice on AWS Lambda, ECS Fargate with FastAPI REST endpoints, Sentry error tracking, and Grafana/Prometheus observability fully automated

Engineered a dual portal consultancy platform using FastAPI and React with Next.js, Firebase Auth with role-based access for student and admin portals, deployed on GCP Cloud Run with auto-scaling, GitHub CI/CD and GCP Monitoring from prototype to production. Built a custom MCP server exposing college search, German university admission criteria, and student profile context as agent tools, enabling the internal recommendation agent to perform dynamically.

Engineered a production RAG backed by LangChain and GraphRAG, processing structured and unstructured data via ChromaDB and Neo4j hybrid retrieval; served 2,000+ daily queries via FastAPI and AWS S3 with WebSocket streaming and sub-2s latency on ECS Fargate. Automated routing of 50,000+ monthly incidents through an ETL and classification by CNN and LSTM pipeline, achieving 90% accuracy; implemented MLOps pipeline on AWS SQS with drift detection, data quality monitoring, and model retraining reducing manual triage time by 65%. Built SynthDataGen microservice, a synthetic data generator with PostgreSQL persistence, exposing functionality via FastAPI with JWT auth and AWS SQS event streaming; reduced data provisioning time by 40% with full CI/CD automation and Prometheus monitoring.