AI/ML Engineer · Computer Vision · Robotics & Autonomous Systems · Medical AI

MS in AI & Autonomous Systems  |  Islamabad, Pakistan  |  muhammadfahadhassan01@gmail.com

I'm an engineer who spent more than a year running production network infrastructure for 10,000+ users, then pivoted fully into AI. That background, real systems, real uptime requirements, real debugging under pressure, shapes how I build: I care about things that actually work, not just things that train.

Right now I'm focused on computer vision for healthcare, financial intelligence, and autonomous navigation, with deployed projects in each area.

Multi-modal financial document intelligence pipeline. ColPali v1.2 indexes PDF pages as image patch embeddings (no OCR) — MaxSim retrieval finds the right page without destroying table structure. Qwen2-VL-7B-Instruct reads tables natively and extracts structured JSON. An agentic risk engine autonomously applies an 8-point financial analyst checklist. Validated on Apple Inc.'s real FY2023 10-K (107 pages, SEC EDGAR): correctly extracted $383B revenue, $97B net income, $6.13 EPS, and flagged the 2.8% revenue decline.

ColPali Qwen2-VL-7B PyTorch BitsAndBytes PyMuPDF SEC EDGAR Dual T4 GPU

Automated psoriasis plaque segmentation and PASI-inspired severity scoring from dermoscopic images. U-Net with EfficientNet-B3 encoder achieves Dice 0.9479 on validation. Severity classifier takes both visual features and computed coverage % as inputs, designed to mirror how dermatologists actually assess severity. Deployed on Hugging Face Spaces with input validation, heatmap overlay, and a PASI-proxy scoring breakdown.

PyTorch segmentation-models-pytorch EfficientNet OpenCV Streamlit ISIC 2018 HAM10000

Procedural ambient synthesis engine that maps natural language to DSP parameters. A custom keyword tokenizer scores text against 80 acoustic clusters to drive a 12-dimension synthesis vector. The system generates audio via a 5-layer Web Audio graph featuring a 25-oscillator detuned pad stack and procedural convolution reverb. It runs entirely client-side with no model weights or external APIs. It includes a custom 16-bit PCM encoder for faster than real-time offline WAV export.

Web Audio API JavaScript DSP NLP Tokenization Additive Synthesis OfflineAudioContext HTML5 Canvas

Multimodal medical diagnostic tool combining U-Net skin lesion segmentation (Dice 0.90, Jaccard 0.85) with a RAG-inspired Llama 3.2 3B Instruct conversational layer. Solved catastrophic forgetting during BioGPT fine-tuning by switching to instruction-tuned Llama with a structured medical knowledge base. Deployed with fuzzy query matching for clinical misspelling tolerance.

PyTorch Llama 3.2 PEFT/LoRA BitsAndBytes Streamlit HAM10000

Monocular SLAM navigation pipeline comparing ORB-SLAM3 trajectory against mobile GPS ground truth. 26.3 cm mean alignment error over full test sequences. Camera calibrated with OpenCV chessboard method; Umeyama similarity transformation applied for SLAM-to-UTM coordinate alignment. Full temporal synchronisation via timestamp interpolation.

ORB-SLAM3 C++ OpenCV NumPy SciPy pyproj Ubuntu

Hardware robot with ultrasonic sensor navigation (95% accuracy across 100 sqm), real-time DHT22 environmental monitoring, mobile app remote control, and ML-based decision logic for zone-targeted cleaning. Data accuracy within ±2%.

ROS2 Arduino Sensor Fusion ML decision logic

Before AI, I spent more than a year as an Operations Engineer at an ISP, managing network operations for 10,000+ subscribers across GPON infrastructure, maintaining 99% uptime SLA, and handling escalated Cisco/Juniper configurations and L2/L3 VPN circuits.

That infrastructure experience is why I'm drawn to problems where the gap between "works in a notebook" and "works reliably" actually matters.

Can be contacted at muhammadfahadhassan01@gmail.com