Work

Intelligent LLM routing system that selects the best model for each task. Outperformed frontier models on SWE-bench benchmarks, reduced inference costs by 50–60%, and scaled to 500+ users.

From-scratch implementation of diffusion models: derived the probability theory, built a VAE, then implemented DDPM step by step to understand image generation from first principles.

Empirical comparison of MHA, MQA, and GQA attention mechanisms, measuring perplexity vs KV-cache size trade-offs on TinyStories. GQA achieved 4x cache reduction with minimal quality loss.

From-scratch implementation of the original Transformer architecture from 'Attention is All You Need', building multi-head attention, positional encoding, and training pipelines.

Machine learning-based threat detection system that identifies and classifies cyber attacks using network traffic analysis.

Research paper on the Mixture of Models (MoM) framework for intelligent LLM routing and model selection strategies.

Framework that lets LLMs analyze git repositories through an iterative REPL loop, exploring code incrementally instead of loading entire codebases into context.

Cross-domain idea generator that solves problems by mapping structural relationships across unrelated fields using analogical reasoning, not surface-level keyword matching.

AI investment assessment tool that evaluates whether an AI project is worth building, with three-year cost projections and industry benchmarks.

AI-powered lead generation tool for pharmaceutical sales teams. Discovers medical professionals across Algeria's 58 wilayas using Google Maps API, with LLM-driven search refinement and quality scoring.

Voice-enabled AI assistant for medical sales reps. Features push-to-talk, real-time transcription, and PubMed integration for citing studies during conversations.

Ad-supported AI chat application offering free access to Claude-level AI, funded by contextual advertising via Thrad's network.

Interactive 3D data visualization platform that turns CSV datasets into particle clouds, controlled by hand gestures via webcam using MediaPipe.

Research replication testing whether prompt repetition improves small language models' accuracy, achieving near-doubling on ARC-Challenge benchmarks.

From-scratch implementation of LSTM networks, building every gate and cell state mechanism from first principles with backpropagation through time.

Lexical analyzer and parser for Mini3ADR, a simple language with variables, arithmetic, and control flow, built with Lex/Flex for a compiler construction course.

Character-level text generator using vanilla RNNs trained on Harry Potter, exploring sequence modeling and the vanishing gradient problem in recurrent architectures.