Aakash Madabhushi
I build the systems a research team runs on: the pipeline that feeds training, the harness that measures the result, and the service that puts it behind an API.
Most of my research work is infrastructure rather than modeling. In the SJSU robotics lab I own the path from raw teleoperation data to a scored policy: pre-training quality gates that refuse to hand a corrupted tensor to a training run, a PyTorch and CUDA pipeline with configurable augmentation and 5K-step checkpointing that can resume a 60K-step run after an interruption, Weights & Biases tracking, and a physical evaluation harness, built for a 100-trial protocol, that scores rollouts one at a time, returns the arm to start on its own, guards against stale caches, and logs each failure mode separately. The researchers get a number they can trust and a run they do not have to restart.
The modeling is real too, across four settings: a 52M-parameter Action Chunking Transformer policy scored on physical rollouts, a graph attention network classifying host-pair links for intrusion detection, a 1.57M-parameter audio-video emotion model built from scratch whose cross-attention fusion matched baselines 30 to 50 times its size, and classical models on tabular data. So is the serving end, where the model has to answer inside a latency budget rather than in a notebook: the threat-detection system returns a retrieval-grounded explanation in 0.87 s at the median, and one model runs behind a FastAPI service that rebuilds and redeploys itself on every commit. Underneath all of it I have built the primitives by hand, from a GPT-2 and a byte-pair tokenizer written from scratch to attention-speedup work and a Phi-3 fine-tune.
Every link goes to code or data you can inspect yourself.
Matched models 30 to 50 times its size on a four-class emotion benchmark with a 1.57M-parameter audio-video network trained from scratch, then showed which facial regions and frequency bands drove every prediction.
SJSU deep learning course, team project · Feb to May 2026
Reached 0.90 F1 flagging suspicious host pairs in 27.1 million Zeek flows, by condensing them into a connection graph with Airflow, dbt and Snowflake, classifying links with a graph neural network, and having Llama 3.3 explain each alert.
SJSU capstone, team project · Feb to May 2026
Took a cancer-recurrence model the whole way to production, by wrapping scikit-learn in a FastAPI service, containerizing it, and redeploying to AWS EC2 automatically on every commit through GitHub Actions.
Personal project · Apr 2025
Reconciled 112 years of California field measurements into one trainable table, by resolving units and station naming that changed repeatedly between 1913 and 2025, then predicting a water-quality score with scikit-learn in Streamlit.
SJSU DATA 245, Machine Learning · Nov to Dec 2025
Scored 92% pick-and-place success and 2.49 cm mean placement error over 50 physical rollouts, by training an Action Chunking Transformer on 50 teleoperated demonstrations and scoring it against a protocol locked before evaluation.
Graduate Research Assistant, SJSU Applied Data Science · Mar 2026 to present
Blocks 100% of payment actions until a person approves them, by reading invoice amounts and dates with Llama 3.2 Vision, scoring its own confidence in each field, and gating every write behind a LangGraph human-approval step.
Personal project · Jun 2026
Reached 0.711 F1 predicting flight delays across 9.5 million U.S. flights, by training a Spark MLlib random forest on 15 monthly BTS files and running it inside a Kafka stream so each incoming flight gets a probability within seconds.
SJSU big data course, team project · Dec 2025
Surfaces Wikipedia pages turning contentious in real time, by streaming every English edit through Kafka as it happens and scoring the tone of each edit summary with RoBERTa.
Personal project · Nov 2025
Select a project for the full detail and links.
I'm available from December 2026 and open to relocating. Email me and I'll reply within a day.
Hiring for something else?