Engineering context-aware AI from model fine-tuning to low-latency production.
BS Artificial Intelligence graduate from FAST-NUCES building context-aware AI systems with deterministic safety controls, from LoRA/QLoRA fine-tuning and vector retrieval to deployment with FastAPI and Supabase.
Featured Engineering Projects
Production-grade systems covering LLM orchestration, hybrid RAG, edge computer vision, and physical dynamics sensor fusion.
OmniDrive AI (Final Year Project)
Edge AI Computer Vision Diagnostics & Physical Dynamics Telemetry Platform
An automotive diagnostic and safety platform combining edge computer vision (YOLO11-Large), physical dynamics sensor fusion via a 2-state discrete Kalman filter, context-grounded RAG assistance, and a transactional multi-role marketplace.
Verified Empirical Metrics
Pipeline Architecture
System Modules
Edge Computer Vision Core (YOLO11-Large)
- -Trained YOLO11-Large champion model on 26,820 images across 50 car-part classes, reaching 96.2% holdout test accuracy (99.34% validation top-1).
- -Optimized inference runtime on CPU constrained under 280MB RAM using ONNX / PyTorch single-threaded execution (~90–110ms latency).
- -On-device pre-compression pipeline in Flutter cutting mobile camera payload by 99.5% (<150ms network dispatch).
Physical Dynamics Telemetry (2-State Kalman Filter)
- -Custom 2-state discrete-time Kalman filter fusing 50 Hz linear acceleration with 1–10 Hz GNSS Doppler fixes.
- -Zero-Velocity Updates (ZUPT) and stationary gravity isolation preventing resting drift under heavy vehicle vibration.
- -Sub-sample linear milestone interpolation calculating millimeter-accurate 0–60, 0–100 km/h, and quarter-mile acceleration benchmarks.
RAG Diagnostic Assistant (Supabase pgvector)
- -// TODO(ammar): pgvector RAG module is verified in repo (api/main.py), add to CV if desired.
- -Context-grounded retrieval-augmented generation engine indexing technical automotive documentation using pgvector (768-dim embeddings).
- -Synthesizes verified mechanical repair workflows via Google Gemini with multi-model fallback and strict XML prompt isolation.
Transactional Multi-Role Marketplace & Security
- -Supabase PostgreSQL managing row-level security (RLS) across 4 roles: Customer, Vendor, Rider, and Admin with TOTP MFA.
- -PostgreSQL row-level write locks (`SELECT ... FOR UPDATE`) preventing race conditions on rider order claims and inventory decrements.
- -Cross-platform Flutter application (Clean Architecture + Riverpod) rendering real-time telemetry HUDs and interactive diagnostics.
OmniDrive Signal-Flow & Edge Processing Pipeline
Recursively denoises sensor noise and suppresses accelerometer drift using predictive covariance updates.
class KalmanFilter1D:
def __init__(self, q=1e-4, r=1e-2):
self.q, self.r = q, r # Process & Measurement variance
self.x, self.p = 0.0, 1.0
def update(self, z: float) -> float:
# Predict & Correct cycle
self.p += self.q
k = self.p / (self.p + self.r) # Kalman Gain
self.x += k * (z - self.x)
self.p *= (1.0 - k)
return self.xWhatsApp Digital Twin — Fine-Tuned LLM & Hybrid RAG
4-Bit QLoRA Qwen 2.5 7B & Supabase pgvector HNSW
Architected a bilingual conversational AI pairing 4-bit QLoRA fine-tuning of Qwen 2.5 7B for stylistic alignment with a hybrid RAG engine for factual grounding. Features a deterministic PII/sensitive-data filtering guardrail and episodic memory retrieval from an HNSW-indexed Supabase pgvector database.
AI Job Application Agent
Autonomous Multi-Agent Workflow with Human-in-the-Loop Review
An autonomous agentic workflow that scrapes job listings, evaluates candidate eligibility using Groq LLaMA 3.3 70B, prepares tailored application responses, automates form-filling via Playwright, and logs outcomes to Google Sheets.
Serene — Mental Health & Crisis Triage AI
DistilRoBERTa Emotion Detection + LoRA GPT-Neo-125M & FAISS
An NLP pipeline pairing DistilRoBERTa for real-time emotion classification with a LoRA-fine-tuned GPT-Neo-125M model for empathetic text generation. Features FAISS vector retrieval for grounding and a deterministic safety layer that intercepts severe distress signals to escalate to human crisis support.
DocuMind — Context-Aware RAG Pipeline
Document-Grounded Conversational QA with Semantic Chunking & FAISS
A document-grounded conversational RAG pipeline using overlapping semantic chunking, MiniLM embeddings, and FAISS top-k retrieval, paired with sliding-window conversational memory to achieve >85% top-3 chunk relevance.
NewsLens & TicketIQ: NLP Classification Systems
Fine-Tuned BERT & Zero-Shot NLI Classification Engine
Production-oriented NLP pipelines combining fine-tuned BERT for hierarchical news categorization (achieving 94% test accuracy) with BART-large-MNLI for zero-shot and few-shot customer support ticket classification and intent routing.
Ferrari LaFerrari: 3D Engineering & Aerodynamics Showcase
Interactive WebGL Hypercar Experience with 963 CV HY-KERS V12 & Audio Synthesis
A high-fidelity 3D WebGL and React Three Fiber engineering exploration of the Ferrari LaFerrari. Features a high-poly 3D model with custom physically based rendering (PBR) metallic car paint shaders, continuous 3D scroll rotation across 7 technical stages, real-time Web Audio V12 engine acoustics, and a bespoke Maranello configurator.
Rolex Calibre 3235: Haute Horlogerie 3D Deconstruction
Continuous Scroll Movement Deconstruction with 240-Frame Canvas Scrubbing & Atelier Configurator
An ultra-luxury interactive continuous-scroll deconstruction of the legendary Rolex Calibre 3235 mechanical movement. Features a high-performance 240-frame Canvas 2D render loop scrubbed via GSAP ScrollTrigger, Lenis inertial smooth scrolling, mechanical ticking sound design, and an Atelier commission configurator.
Protomotive Car Care Studio
High-Performance Interactive Digital Showcase
High-end detailing studio showcase engineered with Next.js App Router, GSAP timeline choreography, Lenis inertial smooth scrolling, and Tailwind CSS v4.
Research Prototypes & Open-Source Experiments
Construction Helmet Detection
Edge computer vision model trained on YOLOv5 for construction safety PPE compliance.
Mesh Shift Visualizer
Interactive WebGL visualization for complex 3D mesh topological coordinate transforms.
Clinical Heart Disease Prediction
Supervised classification pipeline benchmarked across Random Forest and XGBoost ensembles.
Smart Exam Sitting Plan Optimizer
Automated exam hall seating allocation using K-Means clustering to prevent malpractice.
Engineering Experience
Professional engineering internships and academic teaching appointments at FAST-NUCES.
AI/ML Engineering Intern
- •Delivered 5 applied AI systems in 6 weeks (Serene, DocuMind, NewsLens, TicketIQ, Luxe Estate) spanning NLP classification, LLM orchestration, and RAG, shipped as containerized inference services.
- •Built deterministic crisis-safety controls for the Serene health platform that intercept severe-distress signals and override the LLM with safe fallback responses.
AI Developer Intern
- •Shipped production AI workflows using React, Next.js, and the Gemini API, deploying consumer-facing applications on Vercel with event-driven automation via n8n.
- •Designed Supabase persistence layers and validated structured LLM outputs (JSON validation) before feeding them into real-time front-end state management.
Teaching Assistant — Programming Fundamentals
- •Mentored 60+ students in C++ memory management, algorithm debugging, and software architecture.
Bachelor of Science in Artificial Intelligence
- •Graduated with specialization in Machine Learning, Deep Learning, Computer Vision, and Applied NLP.
- •Final Year Project: OmniDrive AI — real-time edge computer vision (YOLO11-Large), 2-state Kalman-filter sensor fusion, and RAG diagnostic platform.
Production Client Systems
Complete production software engineered and delivered directly to business owners, verified in active daily use.
Fine Burger & Fast Food
Custom 3-Portal Direct Ordering, Kitchen KDS & Courier Dispatch Ecosystem
Fine Burger & Fast Food (Est. 1981, Lahore) needed to eliminate high aggregator platform commissions (Foodpanda). I designed, engineered, and delivered a complete direct ordering ecosystem: customer web ordering, a real-time kitchen display system (KDS), and an in-house courier dispatch application.
Eliminated 3rd-party commissions on direct orders, reduced kitchen prep time to under 12 minutes, and automated dispatch across in-house couriers with interactive address map pinning.
Delivered Modules
Progressive web application with dynamic combo selection, address map pinning with reverse geocoding, cart state persistence, and live order tracking.
“Ammar built our entire direct ordering platform from the ground up. Before this, third-party delivery apps were eating 25% to 30% of every order. With our own customer ordering site, kitchen display, and rider app, we have full control over our orders, delivery timing is faster, and our repeat customers love ordering directly from us. He delivered exactly what our business needed.”
Grab Torque
Autonomous AI Software Engineering & Elite Freight Logistics Digital Presence
Engineered the production digital platform for Grab Torque — an enterprise operating two core divisions: custom autonomous AI software development and nationwide freight logistics with wholesale broker load board subscriptions.
Built a clear, high-performance web experience that communicates both technical AI capabilities and freight dispatch operations to prospective enterprise clients and carrier partners.
Delivered Modules
Interactive 4-stage document ingestion simulation (OCR, multi-agent reasoning, database updates), live JSON payload inspector, and 3D visualizer.
“Ammar took complete ownership of our web platform from day one. We needed our site to clearly separate our autonomous AI software services from our freight dispatch business without confusing visitors. Ammar delivered a fast, ultra-smooth 3D experience with custom interactive pipelines ahead of schedule. He's proactive, technical, and solved complex frontend challenges without needing hand-holding.”
Technical Toolkit
Core frameworks and technologies used across machine learning research, inference backends, and full-stack delivery. Highlighted badges indicate core production competencies.
LLM & RAG Orchestration
- LangChain
- pgvector (HNSW)
- FAISS
- LoRA / QLoRA (PEFT)
- MiniLM Embeddings
- Qwen 2.5
- LLaMA 3
- DistilRoBERTa
- Agentic Workflows
Machine Learning & Vision
- PyTorch
- Hugging Face
- YOLO11
- ONNX Runtime
- OpenCV
- Sensor Fusion (Kalman)
- Scikit-learn
- XGBoost
Backend & Data Engineering
- FastAPI
- Python
- PostgreSQL
- Supabase
- Docker
- REST APIs & WebSockets
- SQL
- n8n Automation
- Playwright
Frontend & Mobile Engineering
- Flutter (Clean Arch / Riverpod)
- React
- Next.js
- TypeScript / JavaScript
- Tailwind CSS
- Dart
- Three.js / WebGL
- GSAP
Core Languages & Tooling
- Python
- C++
- Dart
- TypeScript
- SQL
- Git & GitHub Actions
- Linux / Bash
- Vercel
Building systems where the machine learning model is only one piece of the puzzle.
I am an AI/ML Engineer and BS Artificial Intelligence graduate from FAST-NUCES specializing in LLM orchestration, Retrieval-Augmented Generation (RAG), and applied NLP. I build context-aware AI systems with deterministic safety controls—from 4-bit LoRA/QLoRA fine-tuning and dense vector retrieval to low-latency deployment with FastAPI and Supabase.
Alongside language models, my foundation covers edge computer vision (YOLO11), physical dynamics sensor fusion (Kalman filters), and full-stack product delivery with Next.js and Flutter. Rather than building fragile notebook prototypes, I engineer end-to-end production systems where inference latency, throughput, and retrieval relevance are empirically verified.
Verified Credentials
Formal industry internship certifications in AI/ML engineering and production systems.
AI/ML Engineering Internship Certificate
Delivered 5 applied AI systems spanning NLP classification, LLM orchestration, and RAG.
AI Developer Internship Certificate
Full-stack AI workflows using React, Next.js, Supabase, n8n, and Gemini API on Vercel.