Muhammad Aitazaz Ahsan
I'm an AI/ML EngineerFull-Stack DeveloperSoftware EngineerAI/ML Researcher
I build intelligent systems end to end — from fine-tuned models and research pipelines to the full-stack products that ship them.
Two disciplines, one build habit.
I'm a Software Engineering student at NUST (BE, 6th semester, CGPA 3.81/4.00) who works across the full stack of modern AI products — training and explaining the models, and building the web platforms that put them in front of people.
My research leans into the unglamorous parts of ML that matter most: handling a 33:1 class imbalance in system logs, squeezing signal out of audio features with mutual information estimators, making a model's decisions explainable instead of a black box. My engineering side leans into shipping — Next.js, Node, n8n automation, and real users.
Currently exploring agentic AI systems, RAG pipelines, and workflow automation, while interning across computer-vision/identity verification, telecom infrastructure, AI-enhanced web development, and early-stage startups.
Where the work actually happened.
Four internships, four very different environments — computer vision, telecom infrastructure, AI-driven web development, and a fast-moving startup. Tap a card to expand it.
Research & Development Intern
truID · NSTP, Islamabad
R&D Internship
Research & Development Intern
truID · NSTP, Islamabad
R&D intern working on the computer-vision models behind truID's identity-verification platform, plus a new cross-platform seller-verification product.
- Worked across several AI/ML models, with a focus on computer vision model performance
- Found and fixed a critical anti-spoofing bug: high-quality printed photos of faces and palms were being misclassified as real instead of flagged as fake presentation attacks
- Built a hybrid RAG assistant (BM25 + vector search + Gemini Vision) to answer product questions — see the Multi-Modal Hybrid RAG project below
- Contributed to a new product line giving sellers on different e-commerce platforms verified identity checks
Web Developer Intern
Khuda Hafiz (Startup) · Islamabad
Jul 2025 — Aug 2025
Web Developer Intern
Khuda Hafiz (Startup) · Islamabad
Full-stack intern contributing to the company's website and app in a fast-paced startup environment.
- Contributed end-to-end to the company's website and app development
- Implemented new features and improved the UI across customer-facing surfaces
- Maintained backend services alongside a small, fast-moving team
AI Enhanced Full Stack Web Developer Intern
Nexium Ltd · Remote
Jun 2025 — Jul 2025
AI Enhanced Full Stack Web Developer Intern
Nexium Ltd · Remote
Built several AI-powered web applications using Next.js and n8n workflow automation.
- Shipped an AI Blog Summarizer, a Recipe Generator, and a Quote Generator
- Designed n8n automation pipelines to orchestrate AI workflows behind each app
- Sharpened full-stack development and third-party API integration skills
IT Intern
Special Communication Organization · Rawalpindi
Telecom Internship
IT Intern
Special Communication Organization · Rawalpindi
Hands-on exposure to telecom infrastructure and IT operations at a national communications body.
- Worked across media transmission, billing systems, and optical fiber networks
- Exposure to IT & R&D processes, customer care, and call center operations
- Studied CDMA technology and core communication infrastructure
The stack behind the work.
Split across two disciplines by design — the ML side that trains and explains models, and the engineering side that ships them.
AI / Machine Learning
Models, training pipelines, and explainability
Frontend
Interfaces that feel fast and intentional
Backend
APIs and services that hold up under load
Data & Infra
Where the state actually lives
AI Automation
Orchestrating agentic and no-code workflows
Languages & Tools
The everyday toolbox
Selected builds.
18 projects across research, agentic AI, and full-stack products — pulled from the résumé and cross-checked against the real commit history on GitHub.
AI-Powered HDFS Log Anomaly Detector
A fine-tuned DistilBERT pipeline that reads distributed system logs and flags anomalies humans would miss.
- Fine-tuned DistilBERT to model complex temporal and sequential patterns in HDFS log sequences
- Handled a severe 33:1 class imbalance with weighted loss optimization instead of naive resampling
- Achieved a 96.5% F1 score on anomaly detection
- Implemented token-level explainability so the model highlights the exact tokens driving each detected threat
Music Similarity & Recommendation Engine
A recommendation engine built on Mutual Information theory instead of standard collaborative filtering.
- Built on the MagnaTagATune dataset using advanced MI estimators — KSG and KDE
- Optimized spectral audio features with PCA dimensionality reduction
- Ran extensive Grid Search and Leave-One-Out validation to tune the estimators
- Reached a peak benchmark accuracy of 86.98%
Hyperspectral Structured CV
Benchmarks whether structured spatial-neighborhood models actually beat per-pixel classification on hyperspectral imagery — and exposes a common evaluation flaw when they seem to.
- Compared a per-pixel PixelMLP baseline against a spatial PatchCNN3D on the Indian Pines dataset
- Under a naive random split, PatchCNN3D looked 29.6 points better than it should — inflated by spatial data leakage between overlapping train/test patches
Transformer From Scratch vs. DistilBERT
Built a Transformer encoder from raw PyTorch — no pretraining — and benchmarked it head-to-head against fine-tuned DistilBERT on news classification.
- Implemented tokenization, sinusoidal positional encoding, multi-head attention, and stacked encoder layers from scratch
- Trained on AG News (4 classes, 120K examples) with random initialization only
LLM Hardware-in-the-Loop Test Agent
A LangGraph ReAct agent that reads embedded firmware source and docs, writes its own hardware-in-the-loop test suites, runs them, and measures how much it missed against a human baseline.
- ReAct agent loop built on LangGraph, backed by Claude or GPT-4 with Pydantic-validated test schemas
- Tested against a simulated thermostat with a safety-critical latching shutdown — deliberately under-documented to see if the agent finds it via source inspection
Multi-Modal Hybrid RAG Assistant
A conversational RAG assistant over scraped web content, fusing keyword and vector search with multimodal image understanding.
- Hybrid retrieval: BM25 keyword search fused with ChromaDB vector similarity via Reciprocal Rank Fusion
- Gemini Vision converts images into searchable text so the retriever covers non-text content too
Git Whisperer
Paste a GitHub URL, get a searchable, chattable knowledge base of the whole codebase — semantic search, architecture diagrams, and LLM-assisted bug triage included.
- Tree-sitter AST parsing feeds a hybrid pgvector + trigram search (RRF-fused) over the entire repo
- Zero-cost client-side embeddings via transformers.js, with Groq/Llama and Gemini 2.5 Flash for chat and large-context tasks
AQI Predictor — Islamabad
A serverless MLOps pipeline that forecasts Islamabad's air quality three days out, retraining itself on a daily schedule with zero managed servers.
- Hourly ingestion from Open-Meteo and AQICN into a Hopsworks feature store, with 27 engineered features
- Trains five model families per forecast horizon (Ridge, Random Forest, XGBoost, SARIMAX, LSTM) and picks the best independently per day
Hybrid Multimodal Action Recognition
Recognizes human actions by fusing video and smartphone sensor streams in one architecture — two datasets that were never recorded together.
- MobileNetV2 for video features, a 1D CNN for accelerometer/gyroscope sensor streams, fused through a Transformer encoder
- Bridges UCF50 video and UCI Smartphone HAR sensor data via a semantic class mapping, since no synchronized recording existed
Crowd Behaviour Analysis
A CNN-LSTM pipeline for reading crowd behavior patterns from drone and CCTV footage — built as a group research project.
- Combines a CNN for per-frame spatial features with an LSTM over the resulting sequence for temporal crowd dynamics
- Aimed at flagging abnormal crowd movement from aerial and fixed-camera footage
AlphaSite
An autonomous AI platform that generates and deploys full Next.js websites from a single command.
- Engineered a robust n8n backend to orchestrate the entire site lifecycle
- Automated code generation, CI/CD pipelining, DNS configuration, and live deployment
- Designed as an end-to-end agentic workflow with no manual intervention after kickoff
LinkedIn Post & Website Generator
AI-powered n8n automations that turn a prompt into a finished LinkedIn post or a deployed website.
- Streamlined content creation with minimal manual effort
- Chained AI generation steps with publishing/deployment actions
ParkIntel
Intracity smart parking management — real-time slot discovery and navigation.
- Responsive UI built with Next.js and the Google Maps API
- Real-time data handling powered by Supabase
- Secure, one-click authentication via Google Auth
EventEase
A smart event booking and planning platform bridging customers and vendors.
- Dedicated dashboards for bookings, checklists, and service packages
- Real-time messaging between customers and vendors
Kaawish Website
Official platform for Kaawish, a non-profit focused on community uplift and educational support.
- Built the organization's full online presence from the ground up
- Designed to engage supporters and showcase mission-driven programs
Hotel & Restaurant Management System
A desktop management system for hotel and restaurant operations.
- Built with Java and Java Swing
- MySQL-backed persistence layer
Quiz App
A Java-based quiz app that fetches and displays questions dynamically.
- Integrated a trivia API for dynamic question sets
- Built with core Java
Tetris Game
A classic Tetris implementation from first semester.
- Built with C++ and SFML
- First formal game-loop and rendering project
Where the AI/ML Researcher shows up.
Four independent studies, each chasing a hard, unglamorous ML problem — class imbalance, information-theoretic similarity, evaluation leakage, and pretraining-free performance — with measured, defensible results.
AI-Powered HDFS Log Anomaly Detector
A fine-tuned DistilBERT pipeline that reads distributed system logs and flags anomalies humans would miss.
- Fine-tuned DistilBERT to model complex temporal and sequential patterns in HDFS log sequences
- Handled a severe 33:1 class imbalance with weighted loss optimization instead of naive resampling
- Achieved a 96.5% F1 score on anomaly detection
- Implemented token-level explainability so the model highlights the exact tokens driving each detected threat
Music Similarity & Recommendation Engine
A recommendation engine built on Mutual Information theory instead of standard collaborative filtering.
- Built on the MagnaTagATune dataset using advanced MI estimators — KSG and KDE
- Optimized spectral audio features with PCA dimensionality reduction
- Ran extensive Grid Search and Leave-One-Out validation to tune the estimators
- Reached a peak benchmark accuracy of 86.98%
Hyperspectral Structured CV
Benchmarks whether structured spatial-neighborhood models actually beat per-pixel classification on hyperspectral imagery — and exposes a common evaluation flaw when they seem to.
- Compared a per-pixel PixelMLP baseline against a spatial PatchCNN3D on the Indian Pines dataset
- Under a naive random split, PatchCNN3D looked 29.6 points better than it should — inflated by spatial data leakage between overlapping train/test patches
- A strict spatial-block split revealed the true, smaller gain: 70.3% vs 59.8%
- PCA-reduced, per-band standardized features used identically across both architectures for a fair comparison
Transformer From Scratch vs. DistilBERT
Built a Transformer encoder from raw PyTorch — no pretraining — and benchmarked it head-to-head against fine-tuned DistilBERT on news classification.
- Implemented tokenization, sinusoidal positional encoding, multi-head attention, and stacked encoder layers from scratch
- Trained on AG News (4 classes, 120K examples) with random initialization only
- Reached 91.1% test accuracy — 3.5 points behind DistilBERT's 94.6% — at 1/20th the parameter count (3.35M vs 67M)
- Demonstrates the fundamentals hold up even without large-scale pretraining
Straight from GitHub.
Fetched live from github.com/aitazazahsan01, not hand-picked. Whatever ships next shows up here automatically.
0 contributions in the last year
github.com/aitazazahsan01Foundations.
BE Software Engineering
National University of Sciences and Technology (NUST)
Oct 2023 — Oct 2027
Currently in 6th semester with a CGPA of 3.81/4.00, building a strong foundation in software development, systems design, and emerging technologies.
Award of Excellence
Regional Academic Distinction — 2nd Position
Certificates.
Volunteer Work.
What people say after working with me.
“Aitazaz combines strong technical skill with a real eye for user-friendly design — the platform he built for us was polished and easy for our team to adopt.”
Sarah Shafqat
CEO, DreamzSsoft
“A quick learner and highly motivated intern who contributed significantly to the team well beyond what we expected at his level.”
Internship Supervisor
Nexium Ltd.
“Aitazaz has a clear grasp of software design and object-oriented principles, and applies them thoughtfully rather than just by the book.”
Dr. Ayesha Naseer
Faculty, NUST
Let's build something.
Have a role, a research idea, or a product to ship? My inbox is open — I reply personally.