I am Project Lead at Nexus Orbits Pakistan and ML Developer at Zaraat Dost Private Limited, part of AOS Group. My work sits in three areas: geospatial machine learning, private cloud architecture, and multi tenant SaaS products.
Argus is a blueprint for a sovereign private cloud: a complete design for moving workloads off AWS and onto owned hardware, for teams that need data residency, GPU capacity and operational control.
Every component is backed by a decision record that names the options rejected and what they cost, and a validation status page grades each claim as verified, grounded, reasoned or assumed, so the plan says plainly what is still unproven.
| Decisions | 34 architecture decision records, a threat model, runbooks, and a hardware and licensing bill of materials |
| Built so far | Web console prototype with 15 screens, browser RDP and SSH, and 413 passing tests; a self hosted Docker Compose stack; a PowerShell module |
| Platform | Windows Server, Hyper V, Service Fabric, Active Directory, GitOps, zero trust |
| Targets | RPO 15 minutes, RTO 8 hours for full site loss, 99.9% availability |
| Status | Planning stage, v0.12.0. Published to read and discuss, all rights reserved |
| Project | What it does | Built with |
|---|---|---|
| QGIS MCP | QGIS toolbox for open satellite, weather and elevation data, with a permission gated MCP server of 121 tools, an approval queue and an audit log, so AI agents can drive QGIS under human control | Python, PyQGIS, PyQt6, MCP |
| ZD Claude Plugin | Claude Code plugin marketplace that packages a team's geospatial engineering rules into 15 modules, 44 commands, 14 agents and tested hooks. MIT licensed, v7.6.1 | JavaScript, Python, PowerShell |
| AdiCorp HR (live) | Multi tenant HR platform for attendance, leave, payroll, policies and hiring, with every business rule and permission enforced in the database | React, TypeScript, Supabase, PostgreSQL |
| AdiGaze (live) | Resume intelligence platform that turns uploaded resumes into structured candidate records, then matches and ranks them against hiring needs | TypeScript, Supabase Edge Functions, Gemini |
| AdiNox (live) | OTP authenticator and token manager with local persistence and Supabase authentication | React, TypeScript, Supabase |
More: Satellite Imagery Deep Learning, a curated reference list of Earth observation models.
Private: AGIS, a geospatial intelligence platform for satellite analysis and field boundary detection. AdiARC, AI assisted modernization of legacy land records.
Most of my model work is private to Zaraat Dost, so this section lists the architectures rather than repositories.
| Model | Task | Approach |
|---|---|---|
| HRNet W48 encoder with U-Net | Boundary segmentation on 0.3 m imagery | Patch inference with Hanning window blending and hysteresis thresholding, then graph based vector straightening |
| Temporal attention U-Net | Crop mapping from Sentinel-2 | Segmentation over multi date image sequences |
| SAM 2 | Zero shot field boundaries | Tiled inference with seam stitching and cleanup |
| XGBoost | Land use classification | Twelve month spectral and vegetation index features per parcel |
| Isolation Forest | Data cleaning | Removes cloud, shadow and sensor anomalies before classification |
| Gemini | Document extraction and scoring | Resume parsing and candidate ranking in AdiGaze |
| Area | Tools |
|---|---|
| Machine learning | PyTorch, timm, segmentation-models-pytorch, scikit-learn, XGBoost, OpenCV, Albumentations, Hugging Face Hub |
| Geospatial | QGIS, PyQGIS, Google Earth Engine, PostGIS, Rasterio, GeoPandas, Shapely |
| AI engineering | Model Context Protocol, Claude Code plugins, Gemini, Genkit |
| Backend and data | Python, FastAPI, Node.js, .NET, PostgreSQL, Supabase, Firebase |
| Frontend and mobile | TypeScript, React, Next.js, Vite, Tailwind CSS, shadcn/ui, TanStack Query, Expo |
| Infrastructure | Docker, GitHub Actions, Windows Server, Hyper V, PowerShell, Prometheus, Grafana, OpenBao, Vercel, AWS, Google Cloud, Azure |