Piyush Gupta

Machine Learning Engineer · Delhi, India

I ship production computer-vision systems — from aerial intelligence to medical imaging.

Currently building a multi-modal aerial intelligence pipeline at Iris Aerial Innovations — 87 mAP@50 detection serving 2,500 images a day in production testing. Four internships across remote sensing, medical-imaging segmentation, and vision transformers, with depth in multi-modal fusion and inference optimisation under tight compute budgets.

See my work Download resume↓
01 — About

I'm an applied ML engineer and B.Tech student at IIITDM Kurnool with four internships shipping production computer-vision and remote-sensing systems.

My work spans multi-modal aerial detection (RGB, SAR, drone imagery), vision transformers for glaucoma diagnosis, custom U-Nets for terahertz breast-cancer segmentation, and satellite change detection over hundreds of gigabytes of imagery. I care as much about making models deployable — structured pruning, INT8 quantisation, reproducible DVC pipelines — as about raw accuracy. Graduating May 2027, available full-time from June 2027.

Piyush Gupta
Focus
Computer vision, remote sensing, medical imaging, multi-modal fusion
Education
B.Tech, IIITDM Kurnool · 2023 – 2027 · available full-time June 2027
Location
Delhi, India — open to remote & hybrid
Languages
Python, C++, SQL, Bash
Modelling
CNNs (ResNet, EfficientNet, U-Net), Vision Transformers (ViT, DINO, EOMT), YOLO detectors, semantic segmentation, RAG, structured pruning, INT8 quantisation
ML / DL
PyTorch, TensorFlow, Hugging Face, OpenCV, scikit-learn, NumPy, Pandas
Systems
Docker, FastAPI, Git, DVC, Linux, AWS SageMaker, GCP, A/B testing
Highlights
Amazon ML Challenge 2025 — top 500 of 25,000+ teams · Inter-IIIT Hackathon 2026 finalist · Placement Cell Coordinator
02 — Selected Work

Projects with real stakes

GlaucFusion — segmentation-aware glaucoma diagnosis pipeline

GlaucFusion

A bi-directional glaucoma detection pipeline fusing a supervised DINO vision transformer with an encoder-only EOMT — 96% accuracy, +4 points over a GlaucNet baseline across 7 public datasets (30K+ images). Includes a multi-stage OpenCV quality gate (Laplacian sharpness, CLAHE, cup-to-disc ratio) that auto-rejects unusable fundus images, and a controlled CNN-vs-Transformer study across accuracy, generalisation, and compute.

PyTorch · DINO ViT · EOMT · Medical AI

View code on GitHub↗
Terahertz breast-cancer image segmentation with a custom U-Net

Terahertz Breast-Cancer Segmentation

A custom dilated U-Net for tumor segmentation in terahertz medical imagery, trained with a combined BCE + Dice + IoU loss — outperforming a vanilla U-Net on both Dice and IoU under identical training budgets. Includes the full raw-EMF → RGB conversion, augmentation, and heat-map evaluation stack; accepted by Samsung R&D mentors.

TensorFlow · Dilated U-Net · Medical Imaging

View code on GitHub↗
Clinical-RAG — auditable structured summaries from clinical notes

Clinical-RAG

A hospital-grade RAG pipeline that converts unstructured clinical notes into auditable structured summaries. Hierarchical chunking with dual-stage retrieval raises precision over a flat-chunk baseline, and generation is constrained so every emitted field traces back to a source span. Served behind FastAPI in Docker.

RAG · Hierarchical Chunking · FastAPI · Docker

View code on GitHub↗
All projects on GitHub↗
03 — Experience

Where I've worked

  1. Nov 2025 — Present

    ML Engineering Intern · Iris Aerial Innovations

    Architected a 5-stage aerial intelligence pipeline — multi-modal detection, fine-grained vehicle subclassification, temporal change detection, risk scoring, and LLM mission summarisation — serving 2,500 images/day in production testing. Raised 11-class aerial detection to 87 mAP@50 (+31 pts over zero-shot YOLO) and rare-class recall from 66 to 91; unified RGB, SAR, and drone imagery into a 100K+ frame training pipeline; compressed the detector into a 12 GB VRAM budget via structured pruning and INT8 quantisation.

  2. Jul — Oct 2025

    Associate Agent Ops Engineer, ML · AnyTechPros

    Built multi-temporal satellite segmentation and change-detection pipelines over 400+ GB of imagery — 83 IoU and 92 macro-F1 on change masks. Designed SAR–optical fusion preprocessing that improved downstream change F1 by 20 points, and deployed DVC-versioned, reproducible retraining that cut experiment-reproduction time from days to 5 hours.

  3. Jun — Jul 2025

    ML Research Intern · MANIT Bhopal

    Designed GlaucFusion, a bi-directional glaucoma detection pipeline fusing supervised DINO ViT with encoder-only EOMT — 96% accuracy, +4 points over a GlaucNet baseline across 7 public datasets (30K+ images). Ran a controlled CNN-vs-Transformer study and built an OpenCV quality gate harmonising 7 ophthalmology datasets under one schema.

  4. Dec 2024 — Mar 2025

    Research Intern, Samsung PRISM · IIITDM Kurnool

    Designed a custom dilated U-Net for breast-cancer segmentation on terahertz imagery with a combined BCE + Dice + IoU loss, outperforming a vanilla U-Net on both metrics under identical training budgets. Built the full raw-EMF → RGB conversion and evaluation stack; accepted by Samsung R&D mentors.

04 — Research Grounding

Papers & datasets I work from

Papers

Datasets

05 — Writing

Recent posts

Read more on Substack↗
06 — Contact

Open to ML and computer-vision roles.

Internships now, full-time from June 2027. Remote or hybrid, based in Delhi. If you're hiring — or just want to talk about aerial intelligence, medical imaging, or vision models — my inbox is open.

gguptapiyush45@gmail.com Resume↓