Final-year BS Artificial Intelligence @ UMT Lahore 🇵🇰 · Building offline LLM systems, fine-tuning transformers, and shipping AI to the edge.

• 🎓 Final-year BS Artificial Intelligence at the University of Management and Technology (UMT), Lahore
• 🤖 Hybrid specialization: NLP · Deep Learning · Embedded AI — full stack, from fine-tuning BERT to wiring I2S microphones on ESP32-S3
• 🧞 Currently leading GENIE — a fully offline voice AI assistant (Whisper + Mistral 7B + ESP32-S3) for accessibility & smart-environment IoT control
• 🏆 Completed the Advanced AI/ML Engineering Internship at DevelopersHub Corporation — fine-tuned models hitting up to 99.82% accuracy
• 📄 Authored an independent research paper: CNN-LSTM intrusion detection for IoT (99.03% accuracy, CICIDS-2017)
• 🎯 Open to remote AI/ML roles across Europe, Australia & New Zealand
• 📍 Based in Lahore, Pakistan 🇵🇰

"Capabilities once reserved for $50B cloud platforms, running on a $15 board and a laptop."

🧞 GENIE — Offline Voice AI Assistant   Final Year Project · Team Lead

A fully offline voice assistant on ESP32-S3 — no cloud, no API keys, all on-prem.

• 🎤 Hardware: Dual INMP441 I2S mics · MAX98357A class-D amp · 3W speaker · 8MB PSRAM
• 🧠 AI Stack: OpenAI Whisper (STT) → Mistral 7B via Ollama (LLM) → pyttsx3 (TTS)
• ⚡ Latency: 8–15s end-to-end · Cost: ~$22 hardware · Privacy: 100% local
• 🛠️ 10 hard problems solved: dual I2S bus management, PSRAM overflow (480KB buffers), WiFi audio crackle, ghost-voice distortion (22050→16kHz resample), INMP441 bit-shift tuning
• ✅ Phase 1: 16/16 user stories · 10/10 critical bugs resolved
• 🚧 Phase 2 (mostly done): ESP-NOW peer-to-peer ✅ · IoT smart-home control ✅ · Stereo mics ✅ · Wake-word & web dashboard pending

ESP32-S3 Whisper Mistral 7B Ollama Flask I2S Audio PSRAM Offline LLM

Languages

Deep Learning & NLP

LLMs & Agentic AI

Embedded AI / IoT

Data & Tools

Fine-tuned BERT-base-uncased on AG News (120K samples) using HuggingFace Trainer with fp16 mixed precision on RTX 4060. → 94.74% accuracy · 94.75 macro F1 · batch 16 · lr 2e-5 · 3 epochs

Compared zero-shot · few-shot · fine-tuned LLM approaches for ticket classification. → 99.82% accuracy with fine-tuned model · published reproducible comparison report

End-to-end churn prediction on IBM Telco (7K rows) with ColumnTransformer preprocessing + GridSearchCV (5-fold, F1) comparing Logistic Regression vs Random Forest.

Hybrid CNN-LSTM with NSGA-II Jumping-Gene multi-objective feature selection on CICIDS-2017. → 99.03% accuracy · 5× faster training vs MLP/CNN baselines

• 🎖️ Recognition Certificate signed by SVP & Group Head Corporate Affairs, Allied Bank of Pakistan
• 🎓 HEC Pakistan Scholarship — funded Coursera ML & DL specializations
• 🥇 Selected for the Advanced tier of DevelopersHub AI/ML Engineering Internship
• 📜 IBM Generative AI Professional Certificate (Coursera)
• 📜 NVIDIA AI Infrastructure & Operations / Data Centers

I'm actively looking for remote AI/ML internships and full-time roles in Europe, Australia, and New Zealand, and on-site / hybrid roles in Pakistan. Always open to collaborations on NLP, embedded AI, LLMs, or accessibility-focused projects.

📧 aadilmangat@gmail.com  •  📧 armanadil2022@gmail.com  •  🔗 LinkedIn  •  🤗 HuggingFace  •  📍 Lahore, Pakistan

"Ship AI that runs anywhere — even where there's no internet."