Hi, I'm Harish Bejawada
AI/ML Software Engineer
Building agentic AI systems, RAG pipelines and production ML — from prototype to reliable, evaluated deployment.
About Me
Who am I?
I'm an AI/ML Software Engineer with 3+ years of experience designing and deploying agent-based systems using LLMs, tools, memory, and planning for enterprise automation. I work across multi-agent orchestration, RAG pipelines, and human-in-the-loop safety patterns — in Python, LangGraph, LangChain, and AWS Bedrock — shipping production systems with measurable accuracy, reliability, and task success.
Before agents, I spent years on applied NLP and production ML: fine-tuning transformers, building semantic search, and running the MLOps around them. I care about the unglamorous half of AI — evaluation, guardrails, monitoring, and getting a prototype all the way to something people can depend on.
Name: Harish Bejawada
Email: b.harish2727us@gmail.com
Location: United States
Availability: Open to opportunities
My Skills
Agentic AI & LLMs
AI & Machine Learning
MLOps & LLMOps
Vector DBs & Search
Languages & Backend
Cloud & DevOps
Databases
Core Competencies
Featured Projects
Four agents — router, retriever, analyst, critique — answering questions across a PDF corpus with page-level citations. The router decomposes cross-document questions before retrieval; hybrid dense + BM25 search with cross-encoder reranking keeps exact identifiers findable alongside paraphrase. A critique agent verifies every draft against its sources in a fresh context before it ships.
Turns a goal in plain English into a validated plan, then executes it across enterprise APIs with durable human-in-the-loop approval gates. Risk is declared by each connector rather than judged by the agent, so nothing irreversible runs unreviewed. Run state lives in Postgres, so a run parks at a gate, survives a deploy, and resumes — and every step is idempotent.
Multi-step retrieval over a document corpus. Questions are rewritten against the conversation so elliptical follow-ups like “what about enterprise?” resolve into something searchable, then fanned out into several queries, fused, and reranked on whether a passage answers the question — not just whether it is on topic. Streamlit UI with a live retrieval-quality dashboard.
Work Experience
Software Engineer — AI-Driven Financial Intelligence
Moody's Analytics
Designed a multi-agent orchestration system (LangGraph + AWS Bedrock) where planner, retrieval and analyst agents collaborate via tool use to answer regulatory queries, cutting analyst research time by 60%. Built a production RAG pipeline over 10k+ financial documents with hybrid retrieval and grounding checks that reduced hallucinations by 45%, plus human-in-the-loop guardrails and LLM-as-judge evaluation for safe autonomous execution. Shipped low-latency serving (FastAPI, Redis, streaming) that cut end-to-end agent response latency by 55%.
AI/ML Engineer — Applied NLP & Production ML
Cognizant Pvt Ltd.
Fine-tuned BERT and RoBERTa for transaction intent classification across 40+ categories, lifting macro-F1 from 0.78 to 0.92 and retiring a legacy keyword service. Built semantic search over merchant and case records with Sentence-Transformers and ANN indexing (sub-200ms lookups), a LightGBM risk model holding 99.8% precision while cutting false positives 30%, and the serving stack behind them — FastAPI, Redis caching, Docker on ECS at sub-100ms p95. Instrumented drift monitoring and an MLflow retraining workflow that took dataset-to-deployed-model from three weeks to four days.
Junior Data Scientist — Academic Analytics
Gandhi Institute of Technology and Management
Built time-series forecasting models (Prophet, SARIMA) predicting student enrollment and course demand at 87% accuracy, enabling proactive resource planning. Automated ETL across 5+ legacy databases into a unified analytical warehouse, and presented EDA-driven retention insights to academic leadership that shaped two policy changes on at-risk student outreach.