Muhammad Faisal Saeed

Computer Science Student & Data Science Intern specializing in LLM pipelines and Computer Vision.

Overview

A Computer Science student at LUMS with hands-on experience in automating LLM validation and building scalable full-stack applications. He has successfully deployed high-precision machine learning models for geospatial monitoring and real-time edge detection systems.

Pakistan

Experience

Jun 2025 – Aug 2025

INTERNSHIP

Data Science Intern

ILI.DIGITAL

Reduced manual LLM validation effort by ~70% by designing and deploying an automated evaluation pipeline that compared fine-tuned LLaMA model outputs against verified chemical databases, eliminating the need for human spot-checking across 10,000+ daily responses. Accelerated multilingual data ingestion by 5× by building a Python-based PDF translation service using DeepL API and pdfplumber to automatically convert German chemical documents to English, unblocking a key data collection bottleneck. Enabled structured content extraction from 200+ e-learning modules by engineering a SCORM package parser with Python's zipfile and BeautifulSoup libraries, delivering clean JSON outputs consumed downstream by the LLM fine-tuning pipeline.

•Reduced manual LLM validation effort by ~70% by designing and deploying an automated evaluation pipeline that compared fine-tuned LLaMA model outputs against verified chemical databases, eliminating the need for human spot-checking across 10,000+ daily responses

•Accelerated multilingual data ingestion by 5× by building a Python-based PDF translation service using DeepL API and pdfplumber to automatically convert German chemical documents to English, unblocking a key data collection bottleneck

•Enabled structured content extraction from 200+ e-learning modules by engineering a SCORM package parser with Python's zipfile and BeautifulSoup libraries, delivering clean JSON outputs consumed downstream by the LLM fine-tuning pipeline

Key Achievements

→Reduced manual LLM validation effort by ~70%

→Accelerated multilingual data ingestion by 5×

→Enabled structured content extraction from 200+ e-learning modules

LLMLLaMAPythonDeepL APIpdfplumberzipfileBeautifulSoupJSONSCORM

Project Portfolio

ACADEMIC

Present

Punjab Rice Stubble Burning Monitoring System

Lead Developer

Detected province-wide burning events at 89% precision by training an LSTM phenology model on 3 years of Sentinel-2 imagery via Google Earth Engine and deploying a Django + Leaflet dashboard with a REST API. Improved crop-stage F1 by 18% by combining K-Means pseudo-labeling with Random Forest classification to generate growth-stage maps, reducing false-positive fire alerts during non-harvest windows. Delivered real-time geospatial insights by developing an interactive dashboard with Django and Leaflet for district-level fire analytics and air-quality correlation

LSTMSentinel-2Google Earth EngineDjangoLeafletREST APIK-MeansRandom Forest

PERSONAL

Present

Early Smoke and Fire Detection System

Developer

Engineered a real-time smoke and fire detection pipeline using YOLOv5 and MobileNet, purpose-built for edge deployment with high spatial accuracy on constrained hardware. Achieved 92% accuracy on edge hardware using YOLOv5 + MobileNet with INT8 quantization to optimize model size and latency for on-device inference; solved data scarcity by generating 4,000+ synthetic training images through a custom GTA V mod.

YOLOv5MobileNetINT8 quantizationGTA V mod

PERSONAL

Present

CampusCart – Peer to Peer Student Delivery App

Full Stack Developer

Engineered a scalable full-stack marketplace app using the MERN stack, enabling real-time peer-to-peer item and food delivery across campus with live order tracking and buyer-courier coordination. Reduced order coordination time by 40% by building dynamic React.js request forms and Leaflet-based live map routing; hardened the API layer with JWT authentication, bcrypt password hashing, and role-based access control.

MERN stackReact.jsLeafletJWTbcrypt

PERSONAL

Present

3D Reconstruction Mobile Application

Developer

Built a native iPhone camera app that captures images and streams them to a Flask REST API backend, integrating a custom computer vision pipeline directly into a live mobile client. Reconstructed dense 3D meshes from 20 images using an OpenCV structure-from-motion pipeline (SIFT/FLANN feature matching, bundle adjustment, PMVS).

FlaskREST APIOpenCVSIFTFLANNPMVS

PERSONAL

Present

VoxShield – Context-Aware Audio Privacy Redaction System

Developer

Achieved 99% PII redaction accuracy by building an end-to-end pipeline with Whisper-Large ASR and LLM-based NER, mapping redaction decisions to signal-level edits via PyDub and FFmpeg.

Whisper-Large ASRLLMNERPyDubFFmpeg