Madhavan
Balaji.
AI Engineer · MS CS @ Indiana University
I build production AI systems: LLM pipelines, computer vision models, and the ML infrastructure that ships them at scale. Research-backed, engineering-disciplined.
Who I Am
Background
I'm Madhavan Balaji, an AI Engineer with a Master's in Computer Science from Indiana University. I build systems that solve real problems, working across machine learning, computer vision, and applied AI to turn raw data into production-ready software.
From engineering high-accuracy visual retrieval systems and multi-label classifiers to designing end-to-end ML pipelines, I bring both research depth and engineering discipline. I've published findings, filed a patent, and shipped systems that run in production, not just in notebooks.
Beyond building, I care deeply about knowledge transfer. As a Technical Instructor at Indiana University, I helped 30+ students work through complex CS concepts with visual explanations and hands-on mentorship. Today I tutor Statistics for AI at MAC IUI.
Currently
Statistics Tutor at MAC IUI · MS CS, Indiana University, May 2026 · Open to full-time AI/ML roles
Technical Expertise
Comprehensive skills across infrastructure, ML/DL, LLM agents, data operations, and monitoring
Infra & MLOps
Kubernetes
Docker
Jenkins
ArgoCD
AWS
GCP
AirflowML / Deep Learning
Python
PyTorch
TensorFlow
Scikit-learn
C/C++
HuggingFaceLLM & AI Agents
LangGraph
OllamaDataOps & Versioning
DVC
PostgreSQL
Redis
MySQL
Git
DuckDB
PolarsMonitoring & Serving
Prometheus
Grafana
FastAPI
Flask
Streamlit
ReactProfessional Journey
Teaching, tutoring, and applied ML work
Statistics Tutor (Statistics for AI)
MAC IUIStatistical Foundations: Teach statistics for AI, covering the statistical foundations behind machine learning: probability, distributions, hypothesis testing, confidence intervals, and regression.
Concepts to AI Practice: Connect each statistical concept to its AI use, such as regression as a predictive model and hypothesis tests and confidence intervals for judging model results.
Communication: Explain quantitative concepts in clear written and verbal form, adjusting to each student's level.
Technical Instructor & Curriculum Lead
Indiana UniversityEducation Contributions: Designed and delivered weekly technical workshops on Algorithms & Optimization to 30+ students, breaking down complex logic (Dynamic Programming, Graph Theory) into intuitive visual explanations.
Mentorship: Guided students in refactoring and optimal design patterns, mirroring Developer Relations (DevRel) responsibilities by debugging user issues and bridging knowledge gaps.
Community Building: Established a community-first mentorship and peer-review system to drive engagement, demonstrating the ability to build and sustain a technical community.
Teaching is the most efficient form of learning I've encountered. Every whiteboard session, watching a student's face shift from confusion to clarity on dynamic programming, sharpened my own understanding in ways that hours of solo study never could.
Data Analyst Intern
CoincentInference Optimization: Engineered feature extraction pipelines using advanced image processing, reducing latency by 30% for real-time computer vision performance.
Model Development: Developed an end-to-end visual retrieval system achieving 90% accuracy, effectively translating raw visual data into a functional visual search engine.
Recommendations: Built multi-label classifiers for 1,000+ distinct classes, optimizing personalized style recommendations through deep learning retrieval logic.
Featured Work
14 projects spanning autonomous agents, computer vision, data pipelines, and production ML
Sovereign SRE
Detect → Diagnose → Fix → Deploy, all without paging anyone.
Production-grade autonomous SRE platform on 13 AWS services. CloudWatch alarms trigger a LangGraph + CrewAI multi-agent pipeline that diagnoses incidents with codebase-aware RAG (pgvector), generates patches via Amazon Bedrock, and deploys only after human approval via Step Functions + SNS. Live on AWS Amplify.
LLM Eval Harness
Benchmark any LLM across latency, accuracy, hallucination, and cost, in one automated pipeline.
Production-grade benchmarking framework that evaluates OpenAI, Anthropic, and Groq models simultaneously across latency (p50/p95/p99, TTFT), accuracy (semantic similarity, LLM-as-judge), hallucination (NLI cross-encoder), cost, and RAGAS metrics, with Rich terminal output, self-contained HTML reports, and a Streamlit comparison dashboard.
DocuMind
Multi-agent research that reads your documents and writes cited reports, in seconds.
Enterprise-grade multi-agent RAG research platform on AWS. LangGraph orchestrates a 4-agent CrewAI pipeline (Researcher → Analyst → Writer → Critic) that produces fully cited reports using hybrid pgvector + BM25 retrieval with CrossEncoder reranking. Deployed on EC2 behind CloudFront with RDS PostgreSQL, Step Functions, Bedrock Claude Haiku, and Claude Desktop MCP integration.
Project Aegis
A self-healing data pipeline that detects drift and repairs itself, before your models silently fail.
Self-healing data quality system that automatically detects and repairs data drift using Evidently AI and DuckDB streaming. Scales from cloud (200MB) to local (100GB+) with a hybrid architecture and Streamlit dashboard.
Chain-of-Thought Faithfulness
Does your LLM's reasoning actually drive its answers? The data says: probably not.
Research into whether LLM chain-of-thought reasoning actually influences predictions using Negative Preference Optimization. Found fine-tuning on high-quality CoT reduces faithfulness by 56–75%, a data curation problem, not training.
VectorGaze
Natural language safety queries on live video, zero training data required.
Edge AI system detecting workplace safety violations from video using natural language queries, no training data required. Combines CLIP zero-shot embeddings, Moondream2 reasoning, and LanceDB vector search on Apple Silicon.
RiceFlow
Reproducible rice-genomics QC that proves every result it publishes, from laptop to AWS Batch.
CPU-only Nextflow DSL2 pipeline for rice genomic quality control using PLINK 2, BCFtools and SAMtools in Docker. Every run is fingerprinted and checksum-verified, executes locally or on AWS Batch, and publishes to a public read-only dashboard on CloudFront.
StreamCart
Joins what shoppers browsed with what they bought, in one governed lakehouse pipeline.
Retail analytics pipeline unifying live Kafka clickstream with Postgres orders in a Delta Lake medallion architecture (bronze → silver → gold), orchestrated by Airflow, modeled in dbt as a star schema, and served through a Streamlit dashboard.
DentaVision
A Vision Transformer that diagnoses dental conditions at 92.4% accuracy, in under a second.
AI diagnostic system classifying dental conditions (caries, gingivitis) at 92.4% accuracy using a pretrained Vision Transformer. FastAPI backend and React frontend deliver sub-second inference for clinical use.
Hybrid Anime Recommender
A production recommender system with full CI/CD, containerization, and K8s orchestration.
Production recommender combining content-based and collaborative filtering for personalized anime suggestions. Flask API, containerized via Docker and Kubernetes, with Jenkins + ArgoCD CI/CD pipelines.
Multi-Style Image Transfer
Transforms photos into Renaissance masterpieces, Van Gogh canvases, or Impressionist works, on demand.
Transforms photos into art using GAN-based neural style transfer and Stable Diffusion with a custom Van Gogh LoRA adapter fine-tuned on 400 paintings. Supports Renaissance, Impressionism, and more.
OmniBridge
Image captions that actually read the text in the image, not just describe the scene.
Vision-language captioning system built on BLIP-2, fine-tuned with LoRA on TextCaps. Generates accurate captions that incorporate visible text within images for significantly enhanced quality on text-heavy content.
LLM Inference Playground
A hands-on system that exposes what most LLM APIs deliberately hide.
Educational system exposing explicit control over transformer logit space: manual temperature scaling, top-K, and nucleus sampling without HuggingFace defaults. Tracks TTFT and throughput with a FastAPI + React interface.
Daily Tech Intel
3-tier semantic pipeline that scrapes, translates, and summarizes 100+ global tech sources daily.
End-to-end LLM application: a Cloudflare Worker fetches 10 RSS feeds (including Japanese and German sources), passes 100+ articles through Claude for semantic curation into 3 personalized sections, and caches results in Cloudflare KV. On-demand deep analysis with 7-day per-article caching. Built into this portfolio.
Patent & Research
A filed patent for an IoT system that supports elderly care at home
Caremate - Integrated Elderly Support System
A comprehensive IoT-based solution designed to enhance the quality of life for elderly individuals through intelligent health monitoring, emotion detection, and emotional support systems.
Let's Work Together
Hiring for an AI/ML role, or want to talk about a project? Send me a note.
Let's Connect
Email is the fastest way to reach me. I usually reply within a day.
Daily Tech Intel
A daily briefing spanning global AI research, industry moves, and world tech news, automatically filtered to your interests, stack, and research channels. International sources translated to English.