Mar 2026 to presentLab research
Turns 50 human demonstrations into a policy a robot arm runs on its own, scored against a protocol fixed before testing.
Aakash Madabhushi
Robot Learning & Machine Learning Engineer
I build and evaluate learning systems, from robot policies running on real hardware to the data pipelines that feed them.
What I focus on
Robot learning
Manipulation policies trained and evaluated on a real arm, and the simulation around them. Reinforcement learning next: PPO from scratch so far, the arm after.
- Imitation learning
- ACT
- SmolVLA
- LeRobot
- NVIDIA Isaac Sim
- ROS 2
Models on real hardware
Training, fine-tuning and compressing models, then measuring them where they run.
- PyTorch
- CUDA
- LoRA
- INT8 quantization
- ONNX
- Multimodal fusion
Agents & retrieval
Language-model systems that act on real documents, with a check before anything irreversible.
- LangGraph
- RAG
- Vision-language models
- Llama 3.3
- Pinecone
Data pipelines
Batch and streaming pipelines behind analytics and models, 5M+ rows a day in production.
- PySpark
- Airflow
- dbt
- Snowflake
- Kafka
- Databricks
Where I've worked
Research, teaching and production data engineering. Current roles first, then newest first.
Mar 2026 to present
Graduate Research Assistant · SJSU Embodied AI Lab
An ACT policy that succeeds in 46 of 50 real trials on an SO-101 robot arm
5 details for Graduate Research Assistant
- Taught a robot arm to pick up an object and place it on its own, succeeding in 46 of 50 real trials (92%, within 2.49 cm of target), by training a transformer policy that turns camera images into arm motion.
- Brought a broken robot arm back into service by tracing the fault from the camera down to the motors, where a firmware bug silently reassigned motor IDs every time the power was cut.
- Kept a checkpoint with 67% higher action error from being deployed by choosing the version that scored best on held-out tasks, not the lowest training loss. Timing both candidates, 18 ms against 251 ms a decision, ruled the larger one out of real-time control.
- Published 50 robot demonstrations, 49,633 frames, as an open dataset on Hugging Face, with automatic checks that reject a corrupted recording before it can waste a training run.
- Made 100-trial robot evaluations run unattended by building a test harness that resets the scene, recovers from crashes, and logs what went wrong on every trial.
Aug 2026 to present
Instructional Student Assistant, Machine Learning · San José State University
Grading and one-on-ones for 40+ students in graduate machine learning
1 detail for Instructional Student Assistant, Machine Learning
- Supported 40+ students through a graduate machine learning course by grading their work and meeting one-on-one when a lecture did not land.
Jun 2023 to Jun 2024
Software Engineer · Sapaad
5M+ rows a day through production pipelines; dashboards 43% faster
3 details for Software Engineer
- Delivered the data behind every customer's analytics and product recommendations, 5M+ rows a day, by building production pipelines in Python and Spark on Databricks.
- Cut load time on 75+ customer dashboards by 43%, from 12.4 to 7.1 seconds, by profiling the slow Spark queries to see where the time went, rewriting them, and caching repeated results.
- Stopped 14 bad-data incidents before they reached a production model, by putting automatic checks for broken schemas, stale feeds, and failed loads in front of every pipeline.
Sep to Dec 2022
Computer Vision Engineer Intern · Toothlens
Smile previews from one phone photo; 20% more usable patient photos
3 details for Computer Vision Engineer Intern
- Gave orthodontists a realistic preview of a patient's own smile after treatment, from one phone photo, as a working end-to-end service. The clinic had no way to collect before-and-after photos of the same patient, so I trained a CycleGAN, a model that learns the change from two unmatched sets of photos.
- Preserved each patient's face and lighting so only the teeth changed, using CycleGAN's built-in check that an edit followed by its undo must give back the original. It beat Pix2Pix and a VAE on image fidelity over a small hand-built set of matched pairs.
- Increased usable patient photos 20%, across 1,500+, by assembling the before and after training sets and adding a cleanup step that screens out or corrects dark, off-angle, and partly blocked shots.
Sep 2019 to Mar 2020
Research Assistant · Illinois Institute of Technology
Disaster tweets detected live at 85% accuracy, 7 points over baseline
1 detail for Research Assistant
- Detected water-related disasters in live Twitter posts at 85% accuracy, 7 points above the baseline, by collecting tweets as they were posted, turning them into word embeddings, and tuning a classifier.
19 projects, strongest first
Ranked by what each one proves, not by date.
Sep 2026Hackathon
Reads a payments agent's hidden state before each decision and hands risky ones to a safer copy of itself, on one local GPU.
Apr 2026Coursework
Sharpens degraded images in real time on a low-power AI chip, with every design choice made to fit what the chip can run.
Feb to May 2026Team coursework
Reads emotion from face and voice together, matching models 30 to 50 times its size, and shows what changed its mind.
Jun 2026Personal project
Reads an invoice, admits which fields it is unsure about, and cannot pay a cent until a person has approved it.
Spring 2026Coursework
Takes a recruiter from a pile of résumés to a shortlist without slowing the page, and asks before contacting anyone.
Nov to Dec 2025Personal project
Serves the same taxi data to yesterday's dashboards and to live monitoring, from one source, so the two never disagree.
Jul 2026 to presentPersonal project
Tells you where your attention actually went today, and keeps every page you read on your own machine instead of on a server.
Feb to May 2026Team coursework
Finds the suspicious pair of machines among 27.1M network connections, and tells the analyst in plain English why.
Dec 2025Team coursework
Predicts whether your flight will be late, learned from 9.5 million past flights, and scores each new departure in seconds.
Aug 2026 to presentPersonal project
Copies the real table in front of the robot into a simulator, so a policy can practise against the layout it will actually meet.
Nov 2025Personal project
Shows which Wikipedia pages are turning into an argument, right now, by reading the tone of every edit as it lands.
Oct to Nov 2025Personal project
Keeps every revision a price feed makes, so a restated number never silently replaces the one you reported last week.
Apr 2025Personal project
Carries a cancer-recurrence model the whole way from a notebook to a live service that redeploys itself on every commit.
Mar 2026Hackathon
Notices an API has broken, works out why, and fixes it without waking anyone up. Built in five hours at a hackathon.
Nov 2025Personal project
Finds the article you meant rather than the words you typed, by indexing each one by meaning on a nightly schedule.
Nov to Dec 2025Coursework
Makes 112 years of California water data comparable across every unit and station change, then predicts water quality.
Spring 2026Coursework
Lets one part of a review site fail without taking the rest down, because its services talk through events, not direct calls.
Mar 2026Personal project
Gets a robotics simulation stack running on a Mac or a Linux box in one command, instead of an afternoon of driver errors.
Select a project for its figures, stack and links.
No projects match that filter yet.
Built from scratch: 14 earlier projects, language models first
Where I learned the internals rather than the API: a GPT-2 and a byte-pair tokenizer written from first principles, attention made measurably faster, a Phi-3 fine-tune, and then the same treatment for vision models and reinforcement learning. Smaller in scope than the work above, but several have live demos you can try in the browser.
| Project | What it does | Links |
|---|---|---|
| GPT-2 from scratch | Built and trained a GPT-2 language model from first principles | CodeLive demo |
| Transformers from scratch | Built the attention mechanism, then made training measurably faster | Code |
| Tokenizers | Byte-pair encoding tokenizer built for Hindi text | CodeLive demo |
| Phi-3 fine-tuning | Fine-tuned Microsoft's Phi-3 model into a chat assistant | CodeLive demo |
| Multimodal AI system | Combined text, image, and audio inputs in one model | Code |
| Project | What it does | Links |
|---|---|---|
| FastSAM segmentation | Fast image segmentation, deployed as a live demo | CodeLive demo |
| Stable Diffusion | Image generation with style-transfer experiments | Code |
| YOLOv9 object detection | Trained an object detector on a custom dataset | Code |
| UNet & variational autoencoders | Image segmentation and generative models, compared across variants | Code |
| Widget Detector | Turns hand-drawn UI sketches into Flutter components | CodeLive demo |
| Grad-CAM explainability | Visualized which pixels a classifier actually used | Code |
| Project | What it does | Links |
|---|---|---|
| Lightning + Hydra template | A reusable training setup for reproducible experiments | Code |
| CNN foundations | Backpropagation, normalization, and augmentation, worked through by hand | Code |
| PPO reinforcement learning | Agent trained to land a spacecraft in simulation | CodeLive demo |
Two degrees and a certificate
Jan 2025 to Dec 2026 (expected)
M.S. in Applied Data Science · San José State University
Coursework: Big Data Warehousing · Distributed Systems · Machine Learning · Deep Learning · Generative Models & Applications
Aug 2024 to Jul 2025
AI & Machine Learning Operations Graduate Certificate · Indian Institute of Science (IISc)
Aug 2018 to May 2023
B.S. in Artificial Intelligence · Illinois Institute of Technology
Coursework: Data Structures & Algorithms · Operating Systems
What I work with
Every tool here is used in a project above or a role on this page.
- Machine learning & robotics
- PyTorch · CUDA · scikit-learn · pandas · LeRobot · ACT · SmolVLA · Imitation learning · Teleoperation · NVIDIA Isaac Sim · MuJoCo · ROS 2 · OpenCV · Weights & Biases · Hugging Face Hub
- Data & pipelines
- Python · SQL · PySpark · Databricks · Spark SQL · Delta Lake · Parquet · Airflow · dbt · Snowflake · Kafka
- LLM applications & retrieval
- LangGraph · LangChain · Llama 3.3 · Llama 3.2 Vision · Groq · Pinecone · ChromaDB · Sentence Transformers
- Services & delivery
- Docker · FastAPI · GitHub Actions · AWS EC2 · React · TypeScript
- Also worked with
- PyTorch Geometric · Retrieval-augmented generation · Fine-tuning · Gazebo · Redis · MongoDB · MySQL · JavaScript
Beyond the résumé
Background
I'm a Master's student in Applied Data Science at San José State, graduating December 2026, and a Graduate Research Assistant in the Embodied AI Lab, where I train and evaluate manipulation policies on a physical robot arm. I also work as an Instructional Student Assistant for machine learning.
Before grad school I spent a year as a software engineer at Sapaad, building multi-tenant data pipelines that supported analytics for more than 75 customer dashboards at over five million rows a day. Before that I was a computer vision intern at Toothlens, building an unpaired image-to-image pipeline that generated orthodontic previews without the paired patient photos the clinic had no way to collect. My undergraduate degree is a B.S. in Artificial Intelligence from Illinois Tech.
The thread through all of it is that I like problems where you have to build the measurement apparatus before you can tell whether the thing works. Training a policy is the easy half. The harder half is an evaluation harness that survives being interrupted, resets the scene on its own, logs every failure, and tells you honestly that you're at 92% and not at 100%. Right now I'm extending that work into Franka arms, sim-to-real transfer, and vision-language-action models for manipulation.
Teaching
At San José State's AI/ML Club I ran a hands-on LangChain workshop series, teaching students to build applications on top of language models: prompts, chains, and putting the pieces together. Workshop materials
Away from the screen
Most of my spare time goes to a trail, a badminton court, or the gym. Cricket when there are enough of us for a side. And a sudoku when I want a problem that's guaranteed to have exactly one answer, which is not a guarantee the rest of my work offers.