About
Deep learning, inside the clinic.
I'm Digvijay Patil — an AI/ML engineer building systems that integrate directly into the PACS/DICOM workflows radiologists use every day.
At aycan Medical Systems (acquired by the Paratus healthcare IT group), I conceived, designed, deployed, and maintain the entire AI/ML stack — six production systems serving hospitals, clinical-trial organizations, and FDA-regulated workflows. My background pairs an M.S. in Robotics & AI with deep, hands-on expertise in DICOM standards and vendor-neutral imaging architecture.
The throughline is closing the gap between academic AI and the reading room: models that are accurate, run on-premise for HIPAA compliance, and survive contact with real clinical data and real radiologist workflows.
Research interests
Where I'm pointed.
- Foundation-model adaptation for heterogeneous imaging networks
- HIPAA-compliant on-premise inference
- Federated learning across multi-site clinical environments
- LLM-augmented radiology workflow optimization
The path
Career timeline
-
AI/ML Engineer — aycan Medical / Paratus
2023 — Present
-
Software Engineering Intern — Revolutionary Integration
2022 — 2023
-
M.S. Robotics & AI — University at Buffalo
2021 — 2022
-
Engineer — EV / IoT startup
2019 — 2021
-
Software Engineer — SureClaim (medical NLP)
2019 — 2020
Education
-
M.S. Management Information Systems
University at Buffalo (SUNY)
incoming 2026
-
M.S. Engineering Science — Robotics & AI
University at Buffalo (SUNY)
2021 — 2022
-
B.E. Computer Science & Engineering
Visvesvaraya Technological University
2015 — 2019
Stack
-
ML / AI
PyTorch · nnU-Net · Florence-2 · ConvNeXt · CatBoost · Whisper · MediaPipe · Transformers
-
Medical imaging
pydicom · pynetdicom · nibabel · GDCM · NIfTI · DICOM C-FIND/MOVE/STORE · GSPS · MedDream
-
LLM / inference
Ollama · vLLM (dual-GPU) · llama-cpp · EmbeddingGemma · GGUF quantization · on-prem
-
Systems
Python · JavaScript · C++ · Docker · WireGuard · PostgreSQL · Flask · HL7 v2.x · Cloudflare
GPU bench
Optimized across the fleet.
From Apple Silicon MPS to datacenter Pascal — the same model, tuned to whatever's under it. A GGUF quantization ladder (Q8 → Q4 → Q2) and dual-GPU tensor parallelism stretch 7B–70B models onto prosumer hardware.
Apple M1
Metal · MPS
unified memory
dev
Apple M4 Pro
Metal · MPS
unified memory
dev
RTX 5070
CUDA · Blackwell
12 GB GDDR7
work
RTX 5090
CUDA · Blackwell
32 GB GDDR7
personal
Dual Tesla P40
CUDA · Pascal
48 GB · tensor-parallel
server
- MPS backend
- CUDA
- AMP mixed precision
- GGUF quantization
- tensor parallelism
- bitsandbytes
- TorchAO
- ONNX
- flash attention
- KV-cache tuning