ML Engineer · Lahore, Pakistan · Available for hire

Building intelligent
systems that ship.

I design and deploy end-to-end machine learning systems — from raw data to production APIs. Specialized in medical AI, deep learning, and autonomous ML pipelines. I don't just build models. I ship products.

📦 pip install modelscout-ai  ·  Published on PyPI
15+
Projects shipped
94%
Medical AI accuracy
20+
Models in ModelScout
1
PyPI package published
1
Research paper under review

Selected projects

⭐ PyPI Package · Flagship

ModelScout — Autonomous ML Agent

The only tool you need to find the best ML model for any dataset. Upload your data, describe your problem — ModelScout automatically trains 20+ models across tabular, time series, NLP, and computer vision domains, evaluates them with statistical significance testing, and delivers the winner with full proof. Published on PyPI — anyone in the world can install it.

▲ 20+ models · 4 data domains · Statistical significance testing · pip install modelscout-ai
PyTorch TensorFlow Scikit-learn XGBoost LightGBM FastAPI Groq LLM TabPFN DistilBERT PyPI
Voice AI · Offline / Edge RAG

AI Sales Assistant

Fully offline voice-to-voice assistant that answers product questions straight from a company's Excel catalog — no internet, no cloud API cost, runs on CPU-only 8GB RAM hardware. Retrieval-grounded with a strict no-hallucination policy: if it isn't confident, it says so instead of guessing. Knowledge base refreshes from the UI with zero downtime.

▲ 100% offline · CPU-only, 8GB RAM · Refuses to hallucinate
faster-whisper Silero VAD FAISS Qwen2.5-3B (llama.cpp) Piper TTS LangGraph PyQt5
Voice AI · Real-time Agent

AI Receptionist Agent

Real-time conversational voice agent that confirms and reschedules appointments over natural speech — reads and writes a live Google Calendar, logs every outcome to HubSpot CRM, and clarifies ambiguous details (like a time with no AM/PM) instead of guessing. Every call transcript is indexed with hybrid search, so "what did the customer say?" comes back with a cited call ID and date.

▲ Live calendar + CRM integration · Cited call verification
LiveKit Agents Deepgram Groq LLM ElevenLabs Google Calendar API HubSpot CRM Qdrant + BM25 (RRF)
LLM · Legal RAG · Hybrid Search

AI Contract Analyst

Multi-document RAG system for legal contracts, tested on 510 real SEC-filed agreements from the CUAD dataset. Combines BM25 keyword search with vector search via Reciprocal Rank Fusion, keeps conversation memory across follow-up questions, and cites every answer back to the exact source document.

▲ 510 real legal contracts · Hybrid BM25 + vector search · Cited answers
LlamaIndex ChromaDB OpenAI Embeddings Groq LLaMA 3.3-70b FastAPI React + Tailwind
LLM · RAG · Production System

RAG Document Q&A System

Production-grade multi-tenant RAG system — upload PDF, DOCX, or TXT and ask natural-language questions, with every answer cited by document name and page number. Hybrid BM25 + semantic search feeds a cross-encoder reranker for higher precision, ingestion runs async via Celery, and repeat queries return instantly from a Redis cache.

▲ Hybrid search + reranking · JWT multi-tenant auth · Async ingestion
FastAPI Qdrant BAAI/bge-m3 bge-reranker-v2-m3 Celery + Redis Groq LLaMA 3.3-70b
Voice AI · Business Automation

HVAC Voice Agent

Inbound AI voice agent that handles an HVAC business's full customer call flow — answers FAQs, collects customer details, checks real-time technician availability, books the appointment, and emails a confirmation, with zero manual staff intervention. A live dashboard gives staff visibility into every call and booking.

▲ 24/7 inbound calls · Live availability + booking · Zero manual intervention
Vapi GPT-4o mini FastAPI Supabase / PostgreSQL Resend React
LLM · RAG · Documentation Q&A

Streamlit Docs Assistant

RAG-powered chatbot answering Streamlit documentation questions with cited sources. Scraped 273 pages of official docs, built a retrieval pipeline with local embeddings, and grounded every answer in real documentation — with graceful "I don't know" handling for out-of-scope questions.

▲ 273 docs indexed · Cited sources · Zero hallucination on out-of-scope queries
LlamaIndex HuggingFace Embeddings Groq LLM Streamlit BeautifulSoup
FYP · Research · Distinction

SmartLife AI

Prayer-aware smart home automation using stacked LSTM for 17-class activity classification. First ML dataset encoding Islamic prayer times as features. Real-time MQTT backend with zero-downtime HuggingFace deployment. Research paper under review at UET Lahore.

▲ 87.3% accuracy · +9.2% over baseline · Research paper under review
PyTorch LSTM Flask MQTT HuggingFace Firebase
Medical AI · Computer Vision

Brain tumor detection

ResNet18 transfer learning on brain MRI dataset. Deployed as a live REST API — accepts MRI image uploads, returns tumor prediction with confidence score. Data augmentation for small medical dataset generalization.

▲ 94.1% test accuracy · Live REST API · Railway deployment
PyTorch ResNet18 FastAPI Railway Transfer Learning
Medical AI · Explainable AI

Pneumonia detection system

EfficientNet CNN trained on 5,800+ chest X-rays with Grad-CAM explainability — shows exactly which regions triggered the prediction. Critical for clinical trust. Deployed as Streamlit web app with class balancing.

▲ 94% accuracy · 97% recall · Grad-CAM explainability
TensorFlow EfficientNet Grad-CAM Streamlit
LLM · Healthcare · Chatbot

Medical AI Assistant

Conversational AI chatbot that answers general medical questions in real time. Built with Groq API for ultra-fast LLM inference. Clean chat interface with conversation history. Designed with safety disclaimers for responsible AI use in healthcare.

▲ Groq API · Real-time inference · Full chat UI
Groq API LLM Python Streamlit
Computer Vision · NumPy · DSP

NumPy Image Processor

Built-from-scratch image processing engine using pure NumPy — no OpenCV shortcuts. Implements convolution kernels manually: blur, sharpen, edge detection, grayscale, brightness, flip, and rotate. Interactive UI with real-time sliders and drag-and-drop upload.

▲ Pure NumPy kernels · Real-time processing · Interactive UI
NumPy Python Streamlit PIL
Anomaly Detection · Unsupervised

Credit card fraud detection

DBSCAN anomaly detection on 284,807 transactions. K-distance graph for optimal eps selection using formula-based min_samples. Identified 29 suspicious transactions as noise points — demonstrating real-world unsupervised fraud detection.

▲ 284,807 transactions analyzed · 29 fraud cases detected
DBSCAN Scikit-learn Python Pandas
NLP · Text Classification

Consumer complaint classifier

LSTM-based NLP pipeline classifying 5,000 consumer complaints across 5 categories. Complete text preprocessing with tokenization and lemmatization. Outperforms both SVM and standard DNN baselines.

▲ 87% accuracy · Beats SVM (82%) and DNN (85%)
TensorFlow LSTM NLTK TF-IDF
Clustering · Business Intelligence

Mall customer segmentation

KMeans clustering on mall customer dataset identifying 5 distinct personas using Elbow Method. Segments: VIP high-spenders, Extra Spenders, Average, Normal, and Intelligent (high income, low spend). Interactive Streamlit dashboard.

▲ 5 customer segments · Streamlit dashboard
KMeans Scikit-learn Streamlit Matplotlib

Tech stack

🧠
ML / DL Frameworks
PyTorch TensorFlow Keras Scikit-learn XGBoost LightGBM CatBoost
🏗️
Architectures
CNN LSTM Transformers ResNet18 EfficientNet DistilBERT MobileNet ConvNeXt
🚀
Deployment & APIs
FastAPI Flask Docker HuggingFace Spaces Railway Streamlit PyPI
🤖
LLMs & AI Tools
Groq API LlamaIndex RAG Pipelines LangChain LangGraph HuggingFace Vector Embeddings Sentence-Transformers DistilBERT SetFit TabPFN
🎙️
Voice / Realtime AI
LiveKit Agents Deepgram ElevenLabs faster-whisper Piper TTS Silero VAD llama.cpp
📊
Data & Processing
Pandas NumPy OpenCV BeautifulSoup NLTK Matplotlib Seaborn PIL SciPy
🗄️
Backend & Databases
Firebase Firestore MySQL MQTT REST APIs ChromaDB Qdrant Git GitHub Python

About me

I'm a Computer Science graduate from UET Lahore — Pakistan's top engineering university — specializing in machine learning, deep learning, and production AI systems.

I build things that actually work. Every project in my portfolio has real code, real results, and most have live deployments you can test right now. I published ModelScout on PyPI — an autonomous ML agent anyone in the world can install with one command.

My focus is medical AI — applying deep learning where accurate predictions save lives. I've built brain tumor detection and pneumonia detection systems with clinical-grade accuracy and explainability.

I work across the full ML stack — from data cleaning and model training to FastAPI deployment, Docker containerization, and HuggingFace hosting.

UET Lahore CS Graduate 2026 — Pakistan's top engineering institution. FYP awarded Distinction. Research paper under review.

Published PyPI Package — ModelScout (pip install modelscout-ai). Autonomous ML agent covering tabular, time series, NLP, and computer vision.

Multiple live deployments — Railway, HuggingFace Spaces, Streamlit. Real APIs running right now, not just notebooks.

Aspire Leaders Program — Globally selective international leadership initiative across high-achieving international teams.

Let's build something
meaningful together.

Open to full-time ML Engineer roles, freelance projects, and collaborations. Based in Lahore — available immediately.