Syed Muhammad Taha Imam

PROFESSIONAL PROFILE Deep Learning–focused Computer Science undergraduate with hands-on research and industry experience in Transformer architectures, model refinement, and distributed training systems. Proven ability to improve model efficiency through architectural optimization, parameter reduction, and scalable inference pipelines, with experience spanning LLMs, Vision- Language Models, RAG systems, and production-grade MLOps. Actively engaged in research-driven development with a strong inclination toward core DL over black-box model usage. EDUCATION BSCS School of Electrical Engineering and Computer Science , Islamabad , 3.44 (2026) INTERNSHIP EXPERIENCE Epistemy UK 01-Sep-2025 - 01-Dec-2025 – Led the end-to-end development of an AI tutoring platform, architecting a Nest.js backend (20+ end- points, Supabase) and a Next.js frontend. – Architected an event-driven task queue (Redis/BullMQ) to orchestrate multi-agent workflows, ensuring fault tolerance across distributed analysis pipelines. – Enforced strict SWE standards with CI pipelines and pre-commit hooks, achieving and maintaining 90%+ unit test coverage. RapidsAI 01-Sep-2024 - 01-Dec-2024 • Integrated an RAG-based chatbot into a website using Streamlit, enhancing user engagement by 35%. • Optimized chatbot responses through Chain-of-Thought prompting, reducing incorrect responses by 50%. • Cut OpenAI API costs 50% with a multi- model query routing system. • Designed FastAPI endpoint for chatbot responses, enabling user session management and contextual interactions. • Built a Twitter scraper that processed 500+ tweets daily, enabling large-scale sentiment analysis. CogniMind AI 01-Feb-2025 - 01-May-2025 • Deployed a dockerized Apache Airflow instance on a virtual machine, improving workflow automation efficiency by 40%. • Designed 5+ scalable DAGs in Airflow to orchestrate data pipelines, reducing manual intervention by 60%. • Developed CI/CD pipelines with GitHub Actions and Docker, cutting deployment time by 30% • Improved VLM document extraction accuracy by 20% using prompt engineering techniques • Optimized inference pipelines by 10% using parallel and batch processing techniques • Improved retrieval speed and accuracy by 10% using quantization and HNSW parameter tuning Bradbury Lab 01-Apr-2025 - 20-Jan-2026 – Conducted a comprehensive literature review on Transformer topology and PEFT, identifying specific inefficiencies in existing weight-sharing methods. – Proposed a novel layer-merging strategy based on Tucker Decomposition, aiming to reduce parameter count without retraining – Analyzed the mathematical properties of Self-Attention blocks to demonstrate the feasibility of aligning Query (Q) and Key (K) projections for future efficient-by-design architectures. FINAL YEAR PROJECT LiteDoc: Distilling Large Document Models into Efficient Task-Specific Encoders.

Deep Learning–focused Computer Science undergraduate with hands-on research and industry experience in Transformer architectures, model refinement, and distributed training systems. Proven ability to improve model efficiency through architectural optimization, parameter reduction, and scalable inference pipelines, with experience spanning LLMs, Vision- Language Models, RAG systems, and production-grade MLOps. Actively engaged in research-driven development with a strong inclination toward core DL over black-box model usage.