Sunny Shaban Ali.

Final Year Project · R&D

Constructed and released the Close-View Flood Dataset (CVFD) on Kaggle: 7,855 images with pixel-wise semantic segmentation masks across 3 operationally motivated rescue classes, with inter-annotator agreement of Cohen's κ = 0.81. Benchmarked 7 vision architectures on CVFD and identified fine-tuned DINOv3 as optimal at 98.84% accuracy (±0.08) with 21.6M parameters, matching Swin Transformer at one-quarter the parameter count. Built an agentic triage pipeline fusing a CV Agent (DINOv3 + SAM 3) with a zero-shot multilingual LM Agent into a composite priority scorer, validated via Spearman's ρ and Cohen's κ.

Industrial IoT · Dawlance

Real-time OEE / MTTR / MTBF analytics platform, a vendor-neutral replacement for costly legacy industrial monitoring (€60K saved), live on 4 machines at Dawlance and designed to scale to hundreds across multiple PLCs. Per-PLC Tokio daemons on Raspberry Pi read Siemens S7 bytes via snap7-rs and POST hex frames to a Rust (Axum) backend that decodes Q{byte}.{bit} pins through a multi-PLC registry, runs a 4-state machine, and persists to MongoDB with timezone bucketed aggregations. React 18 dashboards stream over Socket.IO with Recharts, guided machine setup, and SAP-linked stoppage classification; secured with JWT, reCAPTCHA, and RBAC; shipped via Docker Compose.

Agentic AI Fitness · MazikGlobal

AI fitness platform anchored by a modular LangChain tool-calling agent (mood tracking, RAG-based Q&A, report generation, profile updates) with CNN-based real-time posture detection in TensorFlow at 94% accuracy. Video-based workout-plan generation via FastAPI and a React progress dashboard; MongoDB-stored evolving user summaries cut chatbot response latency on repeat interactions.

HackOps'25 · 1st Runner-Up

1st Runner-Up at HackOps'25 for an AI skincare analyzer detecting acne and dark circles at 83.8% mAP@50 on a ~2k-image dataset, returning severity scores and matched product/routine recommendations. React + Tailwind + MUI frontend on Vercel with a Dockerized Flask inference service and Supabase, IaC via Terraform; GitHub Actions CI/CD with SonarCloud, GitGuardian, and Dependabot held critical vulnerabilities at zero through release.

Open-Source · 880+ ★

Contributed an open-source screenshot tool to ML4W Hyprland Starter, a popular Arch Linux desktop project (880+ GitHub stars); supports region, fullscreen, interactive, clipboard, and color-picker capture, all triggered via Print Screen shortcuts. Merged 8 keyboard shortcuts (volume, brightness, mute, microphone, WiFi, screen refresh) and a lockscreen display fix for 1920×1080 monitors, across 2 pull requests upstream.

PERN · 2.5K+ users

Built the Coders Cup contest portal in React + Tailwind on Azure Web App with multi-field client validation and Google reCAPTCHA v2, handling 2,500+ users and 900+ team registrations for ACM NUCES. Ranked top-3 on Google for "Coders Cup" keywords with 9,900 page views; shipped via GitHub Actions CI/CD and awarded Special Mention for leading platform development.

Python · OCR

AI-powered OCR tool using Python, PyTesseract, Matplotlib, and PyQt5 to parse 100+ invoice formats at 93% accuracy, reducing manual processing time by 40% through automated extraction and validation.

Interactive Algorithm Visualizer · 27 Algorithms

Interactive visualizer for 27 classic algorithms across sorting, searching, geometry, math — and quantum computing — running one step at a time with synced pseudocode highlighting, plain-English narration of every move, live operation counters, and full timeline control at 1–25 steps per second. Every algorithm is a pure step generator emitting frame snapshots, all driven by a single playback engine. The quantum suite is an exact state-vector simulation: a clickable Bloch sphere, interfering amplitude bars, and Shor's genuine QFT spectrum. A verification harness replays every generator against randomized inputs with 340k+ assertions against reference implementations.