Skip to content

Projects

Systems and research tools I've built — medical imaging pipelines, agent infrastructure, local ML deployment, and the hardware they run on.

WSI Detection Platform

Gigapixel Pathology Inference

A platform for running AI detection models across whole-slide images — tumour-cell and mitotic-figure detection over gigapixel inputs, with tiled inference and region-level result aggregation.

  • Python
  • PyTorch
  • OpenSlide
  • DICOM
  • GPU

Read case study

Encoder or Probe?

Robustness of Pathology Foundation Models

A study of four frozen encoders (Phikon, Phikon-v2, Kaiko ViT-B/16, DINOv2-L) under blur, JPEG compression, and stain shift on NCT-CRC-HE and PatchCamelyon. A linear probe trained on mixed clean and degraded embeddings recovers most of the lost accuracy and reorders the encoders. Published as a Zenodo technical report with code and full results.

  • PyTorch
  • Foundation Models
  • Linear Probing
  • Histopathology

Read the paper Code

Encoder or Head?

Slide-Level Robustness on CAMELYON16

The slide-level follow-up to Encoder or Probe: lymph-node metastasis detection on all 399 CAMELYON16 whole-slide images, streamed, tiled and embedded within a 50 GB disk budget on a 4 GB laptop GPU. Attention-MIL heads trained on mixed clean and degraded tiles recover much of the slide AUC lost to JPEG compression, and tile-level probes show the tumour signal survives in the frozen embeddings. Published as a Zenodo technical report with code and full results.

  • PyTorch
  • Multiple Instance Learning
  • Whole-Slide Imaging
  • Foundation Models

Read the paper Code

NEXUS OS

Bootable AI Operating Environment

A ReAct-style agent orchestrator with skill auto-discovery, a permission model, and a FastAPI server, packaged as a bootable Linux distribution. Extended with vision, RAG, scheduling, and homelab monitoring.

  • FastAPI
  • Linux
  • Agents
  • RAG
  • Vision

Read case study

Local Document-AI Extraction Pipeline

On-Prem OCR + Local LLM

An on-premises pipeline for structured extraction from scanned documents, with canonical item matching and a comparison dashboard — running fully locally on a GPU workstation, no data leaving the machine.

  • pdfplumber
  • Tesseract
  • Ollama
  • Python

Read case study

Distributed Inference Cluster

Heterogeneous Local LLM Serving

A two-node heterogeneous inference cluster (macOS + Windows/WSL) using exo, exploring model sharding and peer discovery for serving large models across commodity hardware.

  • exo
  • Model Sharding
  • macOS
  • WSL

Agricultural Crop-Image Platform — PJTSAU

Professor Jayashankar Telangana State Agricultural University

A mobile application letting farmers upload crop images for expert review by university faculty, designed to curate a labeled agricultural image dataset for training predictive models toward automated crop diagnosis.

  • Mobile
  • Dataset Curation
  • AI Integration

Smart-Farming Weed Detection

Computer Vision on an IoT Field Platform

A computer-vision weed-detection system running on an IoT field platform, classifying crop versus weed from field imagery to help farmers target infestations efficiently.

  • IoT
  • Computer Vision
  • Image Recognition

“orion” Research Homelab

Self-Administered Compute Substrate

A self-administered Linux workstation with RAID storage, zero-trust remote access, and GPU-backed local LLM inference — the compute substrate behind the imaging, document-AI, and agent work above.

  • Linux
  • RAID
  • Zero Trust
  • GPU