Ubaid Ullah

Lead Generative AI Engineer with 2+ years of industrial experience automating, designing, optimizing, and deploying scalable LLM apps using RAG, LangChain/LangGraph, Pinecone/FAISS vector search, prompt engineering, guardrails, and MLOps CI/CD, with hands-on cloud integration across OpenAI/AWS/Azure/Anthropic and RESTful/GraphQL microservices. Completed a 5-month Bytewise Data Science internship (ML/DL/NLP) before delivering 6+ GenAI/RF-sensing research projects as a University of Glasgow Research Assistant (93.23% radar accuracy, 0.89 MRR).

MS Computer Science: 99.34% F1 breast-cancer detection via BioMedCLIP cross-attention + multimodal RAG on 2,006 CDD-CESM mammograms (P@1=0.9950, d=0.994). 3 publications.

Lead Generative AI Engineer - EncoderBytes Private Limited - Peshawar, Pakistan

(2025-10 - 2026-06)

6-month Generative AI Engineer internship under the Pakistan Software Export Board (PSEB), followed by 2 months as a full-time on-site Lead Generative AI Engineer.

• n8n Booking Email Automator – first end-to-end automation project, built independently during the PSEB internship. Solved real-world data-extraction challenges: parsed unstructured booking emails (inconsistent formatting) into clean JSON via IMAP triggers + OpenAI extraction (UK addresses, vehicle types, datetime); handled edge cases, validated data before iCabbi API dispatch; phased rollout (test → semi-auto → full production) with Google Sheets logging for traceability.
• Pharmapedia – AI-powered pharmaceutical content prototype: FastAPI + TinyLlama + Stable Diffusion generating structured drug info (indications, dosage, warnings) and 512×512 illustrations from OpenFDA data; scalable web UI with batch processing for multi-language pharma blogs. [GitHub] [Demo]

Research Assistant (RF Sensing & Telecom) - University of Glasgow - Glasgow, UK

(2024-12 - 2025-07)

• RF-Sensing GPT – multi-modal RAG for 6G: vector+BM25 retrieval + CLIP spectrogram analysis; 13.1% higher faithfulness (0.9617 vs baseline), 93.23% radar accuracy, hierarchical chunking (precision 0.31–0.32, correctness 4.0–5.0), 0.66GB GPU footprint. [GitHub] [Demo]
• TelcoGPT – hybrid embedding framework: 0.89 MRR, 0.94 hit rate (+5–7% vs baseline); term-density scoring → 95% QA accuracy (vs 87%); 93% code-execution success; context 0.92 / accuracy 0.95 / faithfulness 0.92. [GitHub] [Demo]
• RF-Spectrogram VLM – few-shot activity recognition: VLMs with domain prompts, 85.5% subject detection, 82.3% movement, 84.2% activity classification (5-shot); reduces labeled-data dependency for indoor HAR. [GitHub] [Demo]
• Multi-Radar Attention Fusion – adaptive attention: 99.21% accuracy (11 activities) vs single-radar (24GHz: 78.57%, 77GHz: 86.51%, Xethru: 90.48%); fusions: 24+77GHz=96.1%, 24+Xethru=95.8%, 77+Xethru=97.2%. [GitHub]

Data Science Intern - Bytewise Limited - Islamabad, Pakistan

(2024-06 - 2024-10)

• Salmonella Typhi Resistance Prediction (CNN-ANN): fused 28,278 clinical isolates (Lady Reading Hospital, Peshawar; 15-day lab collection) with molecular descriptors; 96.2% accuracy, AUC=0.975 vs. Decision Trees 94.3%; age-dependent patterns (ANOVA F=109.4, p<1e-180). [GitHub]
• Virtual Voice Assistant (Whisper + Llava): real-time accessibility agent – speech-to-text, 4-bit quantized image description, TTS streaming; multilingual Whisper-medium + Gradio UI for visually/auditory impaired users. [GitHub]
• 90 ML/DL tasks over 3 months: basic-to-production progression.

Prompt Engineer - EB Learning Lab - Peshawar, Pakistan

(2023-08 - 2024-06)

• Designed 50+ LLM prompts using few-shot and Chain-of-Thought techniques across training modules.
• Built Python evaluation pipelines with ROUGE/BLEU metrics; accelerated curriculum deployment 2.5×.
• Developed prompt chains deployed to production training platforms for 200+ trainees.

MS - Computer Science - CUSIT (2024 - 2026)

BS - Software Engineering - CUSIT (2019 - 2023)