Portrait of Madhavan Balaji, AI and machine learning engineer
AI Engineer
PyTorchLangGraphFastAPIDockerK8sPostgreSQL
Available now · Open to full-time AI/ML roles

Madhavan
Balaji.

AI Engineer · MS CS @ Indiana University

I build AI you can trust in the real world.

Agents that investigate outages and wait for your approval. Vision models that read the words in a photo. Research that asks whether an AI’s reasoning is honest. Every claim measured, every limit stated.

  • 14 projects, each with its own results
  • Published patent, 2024
  • 29+ students taught
AI agents ✦ Human approval built in ✦ Computer vision ✦ Models trained on a supercomputer ✦ Data pipelines ✦ Results you can verify ✦ Published patent ✦ AI agents ✦ Human approval built in ✦ Computer vision ✦ Models trained on a supercomputer ✦ Data pipelines ✦ Results you can verify ✦ Published patent ✦

Engineer. Researcher. Teacher.

Background

I'm Madhavan Balaji, an AI engineer with a Master's in Computer Science from Indiana University. I'm happiest where research meets engineering: taking an idea that works in a notebook and turning it into something people can actually rely on.

I've built multi-agent systems on AWS that investigate outages and wait for a human to approve the fix, fine-tuned vision-language models on a supercomputer, reproduced and extended a study of how faithfully language models reason, and co-invented a published patent for elderly care.

Teaching keeps me honest. As a Technical Instructor at Indiana University I led weekly algorithms workshops for 30+ students, and today I tutor Statistics for AI at MAC IUI. If I can't explain something simply, I don't understand it well enough yet.

Currently

Statistics Tutor at MAC IUI · MS CS, Indiana University, May 2026 · Open to full-time AI/ML roles

DegreeMS Computer Science
UniversityIndiana University
LocationIndianapolis, IN
FocusAI / ML Engineering
StatusOpen to Full-Time Roles
PatentPublished application, 2024
Teaching30+ students mentored
Projects14 featured

Technical Expertise

Every tool here is one I've used in a real project, from LLM agents and retrieval to data pipelines and the cloud underneath

LLMs & AI Agents

LangGraphLangGraph
LangChainLangChain
LlamaIndex
CrewAI
MCPMCP
Claude APIClaude API
OpenAI API
Amazon BedrockAmazon Bedrock
RAGAS
OllamaOllama

RAG & Vector Search

QdrantQdrant
pgvectorpgvector
ChromaDB
LanceDB
Sentence-Transformers
Hybrid Search (BM25)
Re-ranking

ML, Vision & GenAI

PythonPython
PyTorchPyTorch
TensorFlowTensorFlow
Scikit-learnScikit-learn
HuggingFaceHuggingFace
PEFT / LoRA / QLoRA
OpenCVOpenCV
YOLO
Vision Transformers
CLIP
BLIP-2
Stable Diffusion
GAN
LightGBM

Data Engineering

SQL
PostgreSQLPostgreSQL
Apache SparkApache Spark
Apache KafkaApache Kafka
dbt
Delta Lake
AirflowAirflow
DuckDBDuckDB
pandaspandas
NumPyNumPy
RedisRedis
MySQLMySQL

Cloud & MLOps

AWSAWS
GCPGCP
DockerDocker
KubernetesKubernetes
TerraformTerraform
GitHub ActionsGitHub Actions
JenkinsJenkins
ArgoCDArgoCD
MLflowMLflow
DVCDVC
Cloudflare WorkersCloudflare Workers
HPC / Slurm
GitGit

Apps, APIs & Testing

FastAPIFastAPI
FlaskFlask
StreamlitStreamlit
GradioGradio
ReactReact
Next.jsNext.js
TypeScriptTypeScript
pytestpytest
C/C++C/C++

Where I've built and taught.

Applied machine learning in industry, and the classrooms that sharpened it.

Aug 2026 – Present Indianapolis, IN● Current
StatisticsProbabilityHypothesis TestingRegressionTutoring

Statistics Tutor (Statistics for AI)

MAC IUI
  • Statistical 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.

Aug 2025 – Dec 2025 Indianapolis, IN
AlgorithmsDynamic ProgrammingGraph TheoryMentorshipCurriculum Design

Technical Instructor & Curriculum Lead

Indiana University
  • Workshops: Designed and led weekly workshops on data structures, algorithms and optimization for a cohort of 30+ students, built around live-coding challenges.

  • Leaderboard: Built a competitive-programming leaderboard that tracked weekly and overall class rankings, making progress visible and keeping students engaged.

  • Peer Review: Ran peer-review labs where students compared solutions on efficiency, tricky edge cases and Big-O time and space trade-offs.

Madhavan teaching at an Algorithms workshop, LeetCode problem projected on screen
Weekly Algorithms workshop · Indiana University, 2025

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.

When I started leading weekly workshops on Algorithms & Optimization, I assumed I was there to give knowledge. What I didn't expect was how much I'd receive in return. Students ask the questions that textbooks skip: the "but why does this work?" questions that expose the gaps in your own mental model. Answering those questions live, at the whiteboard, forced me to rebuild my understanding from first principles. I stopped thinking in code and started thinking in structure.

Over dozens of sessions covering Graph Theory, DP, and Greedy algorithms, I developed a discipline I now carry into every engineering context: the ability to explain precisely, not approximately. That precision, earned through teaching, makes me a better collaborator, a cleaner code reviewer, and a sharper problem-solver.

Teaching DP exposed my own gaps

Every time a student asked "but why does the subproblem boundary work here?", I had to rebuild my mental model from scratch, and came out sharper each time.

Graph Theory via whiteboard, not slides

Drawing BFS/DFS live forced me to develop spatial intuition I never got from textbooks. Visual teaching is fundamentally different from visual learning.

Diagnosing mental blockers

A student's confusion is a diagnostic signal. Learning to read it made me a better debugger, interviewer, and communicator on engineering teams.

Weekly prep accelerated my fluency

Preparing workshops 2–3x per week compounded: I solved more LeetCode problems in one semester of teaching than in any prior year of solo study.

Sep 2022 – Nov 2022 Bangalore, India
PythonComputer VisionImage ProcessingDeep LearningData Analysis

Data Analyst Intern

Coincent
  • Faster Features: Cut feature-extraction latency by 30% by profiling bottlenecks and reworking image preprocessing, enabling real-time throughput.

  • Visual Search: Built a CNN-based fashion retrieval pipeline that turned raw product images into a searchable index, with 90% classification accuracy and sub-second queries.

  • Recommendations: Trained a classifier across 1,000+ clothing categories to power downstream recommendations, lifting simulated engagement by 40%.

Things I've built.

14 projects across AI agents, computer vision, data engineering and research. Every number comes from the project’s own results.

2,040 Real IncidentsPython

Sovereign SRE

Finds the cause. Writes the fix. Waits for your yes.

An autonomous on-call engineer built on 14 AWS services. When an alarm fires, AI agents trace the root cause through the codebase, draft a patch and open a pull request. Nothing changes until an engineer approves it by email.

LangGraphCrewAIAmazon BedrockStep Functions
10× Cheaper, Same ScorePython

LLM Eval Harness

Which model should you ship? Now you can measure it.

A benchmarking framework that runs OpenAI, Anthropic and Groq models side by side and scores them on speed, accuracy, hallucination and cost, then puts the trade-offs in one report.

LLM EvaluationRAGASAsync PythonStreamlit
15 ms RetrievalPython

DocuMind

Ask a question. Get a cited research report.

A research assistant that reads your documents and writes a report in which every claim points back to its source. Four AI agents research, analyse, write and critique, on top of hybrid search that retrieves in about 15 milliseconds.

Multi-AgentHybrid RAGLangGraphMCP
50 GB Files, Low MemoryPython

Project Aegis

Bad data, found and fixed before it reaches your models.

A data-quality tool that scans CSV files for missing values, outliers, personal data and semantic drift, then repairs what it safely can. DuckDB sampling lets it handle 50 GB files on a laptop.

DuckDBData QualityPII RedactionStreamlit
Up to 75% Less FaithfulJupyter Notebook

Chain-of-Thought Faithfulness

When an AI explains its reasoning, is it telling the truth?

Research on whether a model’s step-by-step reasoning actually drives its answers. I reproduced a 2025 method across 4 models and 4 benchmarks, then found that training on better-looking reasoning made models 56–75% less faithful.

LLM ResearchInterpretabilityUnlearningHPC
Zero Training LabelsPython

VectorGaze

Describe a safety rule in plain English. It watches for it.

Safety monitoring for video that needs no training data. Write a rule such as “worker without a hard hat”, and VectorGaze finds matching moments with CLIP, then confirms each one with a vision-language model before it raises an alert.

CLIPMoondream2LanceDBZero-Shot

Built in the open.

ContributorOpen-source AI tools

So much of what I’ve learned came from people who shared their work freely. Contributing back is how I say thank you.

I contribute to open-source AI projects because I enjoy solving problems with clarity and care. The best software feels obvious to use, and that rarely happens by accident. It comes from someone noticing the small things and taking the time to get them right. That’s the part I love, and the open-source community is where I get to do it alongside people who care just as much.

  • Start with the userI look for the moments where someone new gets stuck, and fix those first.
  • Explain the whyEvery change comes with its reasoning, its tests, and what I chose to leave alone.
  • Leave it betterSmall, careful improvements that make a tool clearer for the next person.

An invention, on record.

I co-invented a connected system that helps elderly people live safely and independently at home. The Indian Patent Office published it in 2024.

PublishedIndian patent application

CareMate: An Integrated Elderly Support System

An Internet of Things system that helps elderly people live independently and safely. It brings real-time health monitoring, medication management and personalised emotional support together in one connected platform for the person, their family and their carers.

  • Health monitoringContinuous tracking of vital signs, with alerts when readings fall outside safe ranges.
  • Medication managementScheduled reminders and adherence tracking for daily medication.
  • Emotion detectionRecognises signs of distress or low mood so carers can respond early.
  • Companion assistantA conversational assistant that offers everyday support and company.

Let's build something.

Hiring for an AI/ML role, or have a hard problem worth solving? I'd love to hear about it.

Available for opportunities

One email is all it takes.

Tell me what you're working on. I usually reply within a day.

Live project

Daily Tech Intel

A daily briefing of AI research, industry news and global tech, curated around the topics and tools I work with. International sources are translated into English.

For You My Stack Around the World
Open Tech News