Akhil Chintalapati

I find broken systems and fix them.
Then I make sure they actually work in production.

AI Software Engineer  ·  Durham, NC

Want the details? Download résumé →
follow the gold

Kintsugi: the Japanese art of repairing broken ceramics with gold.
Instead of hiding the cracks, you highlight them. The break becomes the most interesting part.

That's how I think about engineering. Find the broken part, understand why it broke, then build something that makes the whole system better than it was before.

Each crack you see below? A problem I worked on.
The gold? The fix that actually shipped.

01
Scale  ·  Autonomy  ·  Impact

Built self-improving AI for
Tesla's supply chain.

The problem: Data scientists were spending 60% of their time doing the same work over and over. Run a query, build a model, answer questions. Repeat tomorrow. Tesla's supply chain couldn't scale like this.

Stakeholders waited days for answers that should've taken minutes. Every decision got bottlenecked because someone had to manually pull data, check it, then explain it to three different teams.

What I built: An AI system that actually learns from what it does. Ask it something Monday, it figures out the answer. Ask something similar Wednesday, it already knows and builds on that. Sub-500ms response time, so decisions happen now instead of next week.

Gave 30+ people their time back. The system now handles work that used to need a data scientist on-call.

Tesla Motors Data Algorithmic Engineer Intern · GenAI & Supply Chain 2025
02
Foundation  ·  Depth  ·  Teaching

Learning AI at Duke,
teaching it to others.

Why Duke: Most AI work is just throwing stuff at the wall. Try this model, doesn't work, try another. Nobody really knows *why* something works.

Duke made me learn the math underneath. Why does this architecture actually converge? Why does that loss function blow up at epoch 15? Now I can predict what'll work before writing code.

Why I teach: I'm a TA for Intelligent Agents. Can't use jargon when students are asking "but why though?" Forces you to actually understand what you're talking about. Makes me write better code.

3.97 GPA while doing this + actual projects. Honestly just really into this stuff.

Duke University M.S. Artificial Intelligence · GPA 3.97 · Teaching Assistant 2024 – 2026
03
Safety  ·  Precision  ·  Alignment

Found safety holes everyone
was shipping to production.

The problem: Your company spends 6 months making a language model refuse harmful requests. Then an engineer fine-tunes it for customer support. Suddenly it'll generate anything you ask. Nobody noticed until I tested it.

This wasn't a one-off. It was systematic. Models that were safe before fine-tuning would happily jailbreak after. The adapter layers were silently undoing months of alignment work, and there were no tests catching it.

What I built: Testing framework that actually measures how much safety you're losing with each fine-tuning run. Now teams can see the tradeoff: "We gained 5% task accuracy but lost 30% safety."

Catch the cost before it ships. Not after users find the jailbreak.

Samsung Research PRISM AI Intern · LLM Safety & Fine-tuning 2023 – 2024
04
Foresight  ·  Signal  ·  Prevention

Predicted failures
before 2am emergency calls.

The problem: Refinery equipment dies at 2am. Production halts. Emergency crews show up. By the time they figure out what broke and fix it, you've burned through $3M in downtime.

Happened every week. The sensors were already screaming 48 hours before the failure, but nobody could separate signal from noise. Just a wall of numbers nobody looked at.

What I built: Models that don't just say "something might break eventually." They say "bearing #3 on pump 7 will fail Thursday around 3pm" with enough confidence that maintenance actually listens.

Maintenance crews fix it Wednesday afternoon during normal hours. Emergency calls dropped 60%. Crisis became just another scheduled task.

Hindustan Petroleum Data Science Intern · Predictive Maintenance 2023
05
Discovery  ·  Invention  ·  Publication

Generated the training data
that didn't exist yet.

The problem: Early oral cancer detection saves lives. Catch it stage 1, you've got a 90% survival rate. Stage 4 drops to 30%. The difference is early detection.

But ML models need thousands of training examples. Rare cancers give you maybe a few dozen. Standard approach: wait 5 years collecting more samples. Patients waiting for treatment don't have 5 years.

What I built: Used GANs and VAEs to generate synthetic training data that looks statistically identical to real cases. The models learn from actual patterns and create more examples of what we're looking for.

Hit 93.5% detection accuracy with the augmented dataset. Got it published in Springer and IEEE.

Research SOAR-ML · Springer CCIS · IEEE Xplore 2023 – 2024
06
Curiosity  ·  Craft  ·  Building

Side projects I built
out of pure annoyance.

Problem 1: Code review bots flag 80% false positives. After a while, developers just ignore everything the bot says. Then real bugs slip through because nobody trusts it anymore. Tool becomes useless.

Problem 2: Got an important meeting at 2pm. Open wardrobe app for suggestions. It recommends the hoodie from yesterday. Has zero awareness of what "important meeting" means or that it's 30°F outside.

What I built: RLHF system that learns from developer feedback - which suggestions were actually useful vs noise. Wardrobe AI that reads my calendar, checks weather, and knows "board presentation" means suit up, not streetwear.

Built both because I got tired of dealing with broken tools. Still working on them.

Personal Projects RLHF Code Review · Multimodal Wardrobe AI Ongoing

The tools of the craft

Python  ·  Rust  ·  TypeScript  ·  C++
PyTorch  ·  JAX  ·  LoRA  ·  RLHF  ·  DPO
FAISS  ·  Pinecone  ·  Hybrid Search  ·  RAG
FastAPI  ·  Django  ·  Redis  ·  PostgreSQL
AWS  ·  Docker  ·  Kubernetes  ·  Airflow
Core ML  ·  ONNX  ·  Quantization  ·  SwiftUI