Aakash Madabhushi
Machine Learning Engineer I train models and build what makes their results trustworthy: clean data going in, honest measurement coming out, and serving that fits a latency budget.
92% pick-and-place success, 50 trials on a real arm at the trained cube position
67% higher action error in the checkpoint kept from deployment, caught by scoring on held-out tasks
30–50× fewer parameters than the emotion models its 74.6% accuracy matched
12× faster training epochs from caching decoded video and audio Took a robot policy from raw recordings to 46 of 50 successful trials on a real arm, with data checks before training and 5K-step checkpoints so a 60K-step run survives an interruption. Kept a checkpoint with 67% higher action error from deployment by choosing checkpoints on held-out task accuracy instead of validation loss, which bottomed out 13K steps too early. Matched emotion models 30 to 50 times larger with a 1.57M-parameter network that reads face and voice together, and made its training 12 times faster by caching decoded video and audio. Ran super-resolution at 8.55 ms an image on a low-power AI chip after compressing the model to 8-bit integers, and flagged suspicious network traffic at 0.90 F1 with a graph neural network. 10 projects in this area Every link goes to code or data you can inspect yourself.
Apr 2026 · Coursework
Super-Resolution on an Edge NPU 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 2026 · Team coursework
Reading Emotion from Voice and Face Reads emotion from face and voice together, matching models 30 to 50 times its size, and shows what changed its mind.
Feb to May 2026 · Team coursework
Explainable Network Threat Detection Finds the suspicious pair of machines among 27.1M network connections, and tells the analyst in plain English why.
Apr 2025 · Personal project
Thyroid Recurrence Predictor Carries a cancer-recurrence model the whole way from a notebook to a live service that redeploys itself on every commit.
Nov to Dec 2025 · Coursework
California Water Quality Prediction Makes 112 years of California water data comparable across every unit and station change, then predicts water quality.
Mar 2026 to present · Lab research
Teaching a Robot Arm to Pick and Place Turns 50 human demonstrations into a policy a robot arm runs on its own, scored against a protocol fixed before testing.
Sep 2026 · Hackathon
Tell: Catching a Payments Agent Being Talked Into Theft Reads a payments agent's hidden state before each decision and hands risky ones to a safer copy of itself, on one local GPU.
Jun 2026 · Personal project
InvoSense: Invoice Processing with a Human Approval Gate Reads an invoice, admits which fields it is unsure about, and cannot pay a cent until a person has approved it.
Dec 2025 · Team coursework
U.S. Flight Delays: Batch Analytics to Real-Time Prediction Predicts whether your flight will be late, learned from 9.5 million past flights, and scores each new departure in seconds.
Nov 2025 · Personal project
Live Wikipedia Edit Sentiment Shows which Wikipedia pages are turning into an argument, right now, by reading the tone of every edit as it lands.
Select a project for its figures, stack and links.
× Super-Resolution on an Edge NPU SJSU course project · Apr 2026 · ML Engineering
Sharpens degraded images in real time on a low-power AI chip, with every design choice made to fit what the chip can run.
Results
8.55 ms per image, running on a low-power AI chip
24.31 dB mean PSNR, over 1 dB above the degraded input
2,217 image pairs trained on, with flip and rotation patch augmentation Built with Python · PyTorch · ONNX · INT8 post-training quantization · Mobilint NPU · U-Net
× Reading Emotion from Voice and Face SJSU deep learning course, team project · Feb to May 2026 · ML Engineering
Reads emotion from face and voice together, matching models 30 to 50 times its size, and shows what changed its mind.
Results
74.6% accuracy across four emotions, on actors never seen in training
1.57M parameters, 30 to 50× fewer than comparable multimodal baselines
12× faster training epochs after caching decoded media, 1,100 s to 88 s Built with Python · PyTorch · CUDA · Cross-attention fusion · Grad-CAM · OpenCV · SpecAugment
× Explainable Network Threat Detection SJSU capstone, team project · Feb to May 2026 · ML Engineering, AI Engineering, Data Engineering
Finds the suspicious pair of machines among 27.1M network connections, and tells the analyst in plain English why.
Results
0.90 F1 flagging suspicious machine pairs (0.94 precision)
27.1M Zeek network flows aggregated into the graph
0.87 s median end-to-end latency per retrieval-grounded explanation Built with PyTorch Geometric · Airflow · dbt · Snowflake · Pinecone · Llama 3.3
× Thyroid Recurrence Predictor Personal project · Apr 2025 · ML Engineering, Software Engineering
Carries a cancer-recurrence model the whole way from a notebook to a live service that redeploys itself on every commit.
Built with Python · scikit-learn · FastAPI · Docker · GitHub Actions · AWS EC2
× California Water Quality Prediction SJSU DATA 245, Machine Learning · Nov to Dec 2025 · ML Engineering
Makes 112 years of California water data comparable across every unit and station change, then predicts water quality.
Built with Python · scikit-learn · pandas · Streamlit
× Teaching a Robot Arm to Pick and Place Graduate Research Assistant, SJSU Embodied AI Lab · Mar 2026 to present · Robot Learning, ML Engineering
Turns 50 human demonstrations into a policy a robot arm runs on its own, scored against a protocol fixed before testing.
Results
92% pick-and-place success over 50 real trials, at the trained position
2.49 cm average placement error on successful trials
50 teleoperated episodes published as an open dataset, 49,633 frames Built with Python · PyTorch · CUDA · LeRobot · ACT · SmolVLA · Weights & Biases · Hugging Face Hub
× Tell: Catching a Payments Agent Being Talked Into Theft SJSU hackathon, run on an HP ZGX Nano · Sep 2026 · AI Engineering, ML Engineering
Reads a payments agent's hidden state before each decision and hands risky ones to a safer copy of itself, on one local GPU.
Results
0.841 AUROC on held-out invoice attacks, 400 records from 50 documents
17 GB GPU memory to serve Qwen3-8B and its adapter on one GB10
76 / 76 clean invoices still paid correctly with routing on Built with Python · PyTorch · CUDA · Qwen3-8B · Hugging Face Transformers · LoRA · NVIDIA GB10 · Tesseract · SQLite
× InvoSense: Invoice Processing with a Human Approval Gate Personal project · Jun 2026 · AI Engineering, ML Engineering
Reads an invoice, admits which fields it is unsure about, and cannot pay a cent until a person has approved it.
Built with Python · LangGraph · Groq · Llama 3.2 Vision · Llama 3.1 · ChromaDB · React
× U.S. Flight Delays: Batch Analytics to Real-Time Prediction SJSU big data course, team project · Dec 2025 · Data Engineering, ML Engineering
Predicts whether your flight will be late, learned from 9.5 million past flights, and scores each new departure in seconds.
Results
0.711 F1 predicting delays, 79.1% accuracy
9.5M flight records processed, 15 monthly BTS files totalling 3.1 GB
0.3% train-minus-test accuracy gap, against 1.6% for a single decision tree
78% of random-forest importance mass held by the top five features Built with Python · PySpark · Spark MLlib · Spark SQL · Kafka · Structured Streaming · Streamlit · Docker
× Live Wikipedia Edit Sentiment Personal project · Nov 2025 · Data Engineering, ML Engineering
Shows which Wikipedia pages are turning into an argument, right now, by reading the tone of every edit as it lands.
Built with Python · Kafka · RoBERTa · Hugging Face
Hiring for machine learning engineering? I'm available from December 2026 and open to relocating. Email is the fastest way to reach me.
Hiring for something else?
