Harshith Chennupati
AI/ML engineer for biomedical signals, medical imaging, and deployable clinical systems.
I build end-to-end ML systems that turn clinical and physiological data into validated, shippable outputs — sleep staging and fragmentation from multi-modal biosignals, ECG respiratory-rate estimation, GPU-accelerated 3D imaging pipelines, and tooling that puts models in the hands of clinicians.
GPA 3.8/4.0 · AI, Biomedical Signal Processing, Medical Imaging, Medical Device Design · Deep Learning, ML for Signal Processing, ML for Medical Applications, Computer Integrated Surgery
GPA 3.76/4.0 · Embedded Systems, Signal Processing, FPGA Design
Wearable sleep staging via PSG-to-IMU distillation, IMU pose estimation, and team MLOps.
- Built a PSG-teacher to IMU-student knowledge distillation pipeline for wearable sleep staging on the DREAMT corpus; benchmarked classical biosignal baselines (Cole-Kripke, vanHees) against transformer-based sequence models with a rigorous evaluation harness across paired biological corpora.
- Developed IMU-based human pose estimation with clustering (K-means, GMM, HDBSCAN) on bicep-mounted sensor streams; authored the ground-truth labeling protocol for posture-class validation.
- Established MLOps practices for the team: experiment tracking, dataset versioning, reproducible training configs, and evaluation dashboards for fast iteration.
Cross-modality 3D synthesis, GPU vision pipelines, multimodal acquisition, and clinician-facing inference tooling.
- Built MAISI-US, a ControlNet-conditioned 3D latent diffusion pipeline adapting NVIDIA's MAISI foundation model to cross-modality ultrasound-to-CT synthesis. First reported adaptation of MAISI to a non-CT input modality; trained only the ControlNet branch on 9,068 paired US-CT volumes for low-cost domain transfer while keeping the VAE and diffusion backbone frozen.
- Designed a multi-metric evaluation protocol (PSNR, mutual information, gradient orientation agreement, LPIPS) against paired ground truth, with parameter sweeps on inference steps, ControlNet scale, and noise factor for reproducible characterization of model behavior.
- Built GPU-accelerated 3D vision pipelines (volumetric resampling, 3D-2D registration, DRR generation) on a CUDA-based imaging library; implemented pose regression with a custom Res-UNet for real-time 2D-in-3D visualization.
- Led multimodal data acquisition integrating Clarius and Alpinion ultrasound probes, Medtronic O-arm CBCT, NDI optical tracking, and scripted Universal Robot (UR3) sweeps to produce a paired calibrated 3D dataset for training and quantitative evaluation.
- Shipped Qt-based tooling for 3D Slicer to orchestrate deep-learning inference, putting research models directly into the hands of clinicians and researchers.
ECG-based respiratory estimation and accelerometer posture classifiers for a wearable monitoring product.
- Designed ECG-based respiratory rate estimation using CEEMDAN, discrete wavelet transforms, and peak detection, achieving 2.5 bpm MAE on the BIDMC dataset.
- Implemented and unit-tested accelerometer-based posture classifiers for a wearable monitoring product.
Course support for CNNs, segmentation, and multimodal registration.
- Designed and graded assignments on CNNs, segmentation networks, and multimodal registration.
- Mentored students on PyTorch implementations, training diagnostics, and evaluation.
ECG biosensor testing and optimization in hospital environments for FDA-approved wireless devices.
- Optimized ECG data collection and testing procedures in hospital environments for FDA-approved wireless biosensors.
- Contributed to A2 ECG biosensor design; gained hands-on exposure to medical-device SoC and embedded workflows.
Sleep-staging and biosignal ML for a clinical wearable — building evaluation pipelines that compare deep sequence models and classical baselines against polysomnography ground truth. Proprietary work; details limited by NDA.
ControlNet-conditioned 3D latent diffusion adapting NVIDIA's MAISI foundation model for ultrasound to CT synthesis using paired cadaver US-CT ROIs, VAE latent pre-encoding, and a multi-metric validation suite (PSNR, SSIM, LPIPS, GOA).
Unified MCP server for federated medical/biosignal dataset discovery — one query across PhysioNet, NSRR, Zenodo, HuggingFace, OpenNeuro, and Kaggle, returning ranked, normalized results for LLM clients.
Implemented 3D registration, calibration, error-modeling, and distortion-correction modules; cut ICP runtime 40% with bounding-volume hierarchies and reduced EM-tracking positional error by 30%.
Trained a multi-task U-Net with residual blocks on the SIIM-ACR pneumothorax dataset; achieved IoU 0.65 and 80% accuracy with a custom class-imbalance loss.
Implemented ISTA-Net and ADMM-Net for accelerated MRI reconstruction from 25% undersampled k-space; extended ADMM-Net with Squeeze-and-Excitation blocks for 1.14× speedup at 24 dB PSNR.
Engineered an at-home diagnostic ML pipeline from smartphone-recorded lung sounds using feature extraction, filtering, and classification on the ICBHI dataset — 95% accuracy on clean data.
Improved noisy-audio classification accuracy from 18% to 51% using MetricGAN+ denoising and Wav2Vec2 features; gained an additional 20% via HuBERT and BERT multimodal fusion on the MELD dataset.
Shipped Qt-based extensions for 3D Slicer that orchestrate deep-learning inference, so researchers and clinicians can run models without leaving the imaging environment.
A signal-processing pipeline that estimates respiratory rate directly from a single-lead ECG, achieving 2.5 bpm MAE on the BIDMC benchmark.