Raahim A Samad Poonawala

Raahim A Samad Poonawala

AI Research Engineer | VLM & LLM Specialist | Full-Stack Startup Architect

Based in: KarachiCurrently: Research Intern (Medical VLM), Tibbling TechnologiesExperience: 3 years

Hire me

CS Graduate from LUMS specializing in **Deep Learning** and **Autonomous Systems**. From architecting **GRPO-based frameworks** for LLM generalization to building **Alzheimer's detection VLMs**, I bridge the gap between high-level research and production-ready engineering. At **onBeacon**, I scaled customer intelligence by integrating **RAG workflows** and **LLM verification** into a full-stack environment. I thrive on solving complex engineering hurdles, whether it's stabilizing training loops for **GPT-like models** from scratch or optimizing satellite imagery pipelines to outperform state-of-the-art benchmarks by **29.39 dB PSNR** (winning the Best Project Award). I am driven by the goal of making AI systems more **traceable, reliable, and impactful** for end-users.

Experience

Research Intern (Medical VLM)

Tibbling TechnologiesinternshipPresent

Summary

Researching multimodal vision-language models for clinical Alzheimer's detection.

What I did

  • Source and reconcile medical imaging datasets, preprocessing them into usable multimodal formats.
  • Fine-tune and evaluate vision-language models for Alzheimer's detection.
  • Establish baselines for model behavior across Brain MRI benchmarks and transition the project into a systematic, reproducible pipeline for clinical evaluation.
VLMBrain MRIMedical ImagingMultimodal LearningFine-tuningPython

Web Developer & Machine Learning Engineer

Independent WorkworkMay 2024

Summary

Leading independent research and development across machine learning, full-stack AI, and financial intelligence.

What I did

  • Developing a RAG-driven financial intelligence platform for PSX equities, integrating real-time market data and AI agents for automated analysis.
  • Built and optimized GPT-like architectures (Wavenet, Bigram) from scratch, resolving vanishing gradient issues to stabilize training loops.
  • Implemented a 96% accurate ViT-based food classifier trained on 100k+ images.
  • Improved platform efficiency by 70% for the Kissan-Dost IoT integration using FastAPI and RAG pipelines.
PyTorchMERN StackThree.jsFastAPIRAG pipelinesORMsIoTViT

AI Software Engineering Intern

onBeaconworkJun 2026 – Aug 2026

Summary

Engineering RAG workflows and full-stack AI integrations for product feedback.

What I did

  • Integrated Granola and Notion through MCP/APIs for automated feedback ingestion.
  • Built RAG workflows linking product-flow and UI audits to customer evidence using OpenAI and Anthropic for high-fidelity verification.
  • Optimized system reliability by migrating core workflows to a PostgreSQL, Redis, and Celery stack while contributing to full-stack development in FastAPI and Next.js.
FastAPIPostgreSQLRedisCeleryNext.jsFlaskOpenAIAnthropicRAGAPIs

Skills

Technical

PythonPython
CNNs
Vision Transformers
PyTorchPyTorch
Reinforcement Learning
Transformers
Curriculum RL
MERN Stack
Q-learning
Policy Gradients
Diffusion Models
scikit-learn
Pandas
Numpy
Three.js
HuggingFace
YOLOv10
SEO
FastAPIFastAPI
SQLSQL
Azure
MATLAB
TypeScriptTypeScript
Next.jsNext.js
ReactReact

Areas of expertise

Leadership
Community Service
Project Management

Projects

Dynamic Programming and Reinforcement Learning (CS6314) - Course Project

Developeracademic
Developed a curriculum-trained autonomous multi-agent driving agent using Python, PyTorch, OpenAI Gym and MetaDrive, outperforming non-curriculum baselines across complex traffic scenarios. Achieved up to ~45% improvement in mean reward and ~22% higher task completion rates, while reducing training time by ~25-30% through staged learning.
PythonPyTorchOpenAI GymMetaDrive

Kahani Project

Research Internprofessional
Built and pre-processed 10k+ entries Urdu dataset for the "Kahani" project using Python-based web scraping and data cleaning scripts.
PythonWeb scraping

ViT-based food classifier

Developerpersonal
Implemented a 96% accurate ViT-based food classifier using PyTorch, trained on 100k+ images across 101 categories.
PyTorchVision Transformers (ViT)

Kissan-Dost platform

Developerpersonal
Optimized the Kissan-Dost platform using FastAPI, RAG pipelines, ORMs, and real-time IoT data integration, improving platform efficiency by 70%.
FastAPIRAG pipelinesORMsIoT

Advanced Topics in Machine Learning (CS-6304) - Course Project

Researcher/Developerpersonal
Proposed RewardMaxxing, a novel GRPO-based framework integrating token-level memorization penalties via the STIM framework. While the project faced an inference database bottleneck that limited final benchmarking, the architecture was designed to improve LLM generalization under distribution shifts.
PythonPyTorchGRPOSTIM framework

Darcheeni

Research Internpersonal
Fine-tuned and benchmarked on 7+ hours of medical data using LoRA and prefix tuning techniques for "Darcheeni."
HuggingFaceLoRAPrefix tuningPython

GPT-like models from scratch

Developerpersonal
Built GPT-like models from scratch (bigram, wavenet, makemore, micrograd), gaining deep architectural insight.
PythonBigramWavenetMakemoreMicrograd

DeployTeach

Machine Learning Engineerpersonal
A privacy-preserving classroom analytics system designed to study teacher pedagogy without relying on cameras or fixed sensing infrastructure. The system combined instructor location/movement tracking with speech diarization to understand where the teacher moved, when they were speaking, and how those patterns related to classroom interaction.
Silero VADsherpa-onnxpyannotePython

DVLM Latent-Reasoning Project

personal
Research-focused vision-language modeling project exploring whether reasoning could be improved through latent/internal representations rather than relying only on explicit chain-of-thought style outputs. The work involved building evaluation pipelines, running ablations across different model/training setups, and experimenting with value-model and PPO-style components to study reasoning quality and generalization.
Vision-Language Modeling (VLM)PPOValue-modelPythonEvaluation pipelines

PSX Financial Intelligence Platform

personal
A personal financial-intelligence platform focused on PSX equities, combining price/volume histories, company fundamentals, and news filings with RAG-driven AI agents for stock analysis.
PythonRAGAI AgentsFinancial AnalyticsReactNext.jsTypeScript

Education

BSc (Hons.) Computer Science

Lahore University of Management ScienceComputer ScienceGPA 3.76Sep 2023 – 2027

Coursework

Internet of ThingsDeep Vision Language ModelsSoftware EngineeringAlgorithmsData Structures

Societies & activities

AIESEC in LUMSSPADESLUMS Religious Society

Recognition

Awards

Best Project Award

WiT SMP 20252025

Awarded for the coarse-to-fine super-resolution system for soil moisture prediction.

Best Delegation Award

CWMUN'192019

Won for Pakistan as a part of the WHO committee.

Dean's Honours List

Lahore University of Management Science

Awarded for a 3.6+ GPA for Freshman and Sophomore year.

Certifications

Fundamentals of Digital Marketing

Google2023

CS50X: Introduction to Artificial Intelligence (and its Neural Networks)

edX2023

AZ900 & AZ104: fundamentals and associate developer of Microsoft Azure

MicrosoftExpires Aug 20252025

Social impact & activities

Business Development Team Leader

AIESEC in LUMSorganizerAug 2024 – Dec 2024

Organized "Smile for LUMS" in partnership with HBL, benefiting 80+ staff members with donations and financial education

Hire RaahimReach out about a role, a contract or a conversation.For recruiters

Raahim's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis