I build AI systems end to end — the pipeline, the model integration, the API, and the interface it all lives in. The through-line in my work is local-first AI: agents, RAG with cited sources, and fine-tuned small models that run where the user lives, not in someone else's cloud.
MS COMPUTER SCIENCE @ STONY BROOK UNIVERSITY (AUG 2024 — MAY 2026) · RESEARCH & STUDENT ASSISTANT · NEW YORK
SCROLL-DRIVEN FILM REELS · BUILT WITH VANILLA HTML/CSS/JS · CLICK TO ENTER — azarkazar.github.io/mohamed-azar-portfolio
 Investor Intelligence Platform Annual reports in, answers out — LLM KPI extraction with confidence-gated retries, plus hybrid keyword + vector RAG search over filings, on a live dashboard. 
|
 Support Ticket Triage — LLM Fine-tuning QLoRA fine-tune of Qwen2.5 / Llama-3.2 classifying support tickets across ~27 intents — 4-bit quantization + LoRA adapters, trained and tracked on Azure ML. 
|
 Earnings-Call Evasiveness Detector Fine-tuned small language model labeling executive answers Direct / Partial / Evasive on human-reviewed earnings-call transcript data.
|
 Blog → Podcast Agent Paste a URL, get a narrated episode — scraped, summarized by a local LLM, and voiced live, end to end, on a two-tier FastAPI + Next.js system.
|
 Customer Support Voice Agent Local RAG support assistant with cited answers, hallucination safeguards, automatic ticket creation on low confidence, and spoken replies. 
|
 GitHub MCP Agent Explore and analyze any repository in natural language — an MCP GitHub server driven by fully local Ollama reasoning. 
|
 AI Breakup Recovery Companion Privacy-first wellness app — four specialized local agents (Therapist, Closure, Routine Planner, Brutal Honesty), OCR chat-screenshot analysis, zero cloud calls.
|
 Athlete Injury Risk Analyzer Biomechanical symmetry metrics → injury-risk classes; Random Forest compared across SMOTE / SMOTEENN and class weighting, with hypothesis testing. 
|
LOCAL-FIRST AI · AGENTS · RAG · LLM FINE-TUNING · END-TO-END SYSTEMS — azarkazar.github.io/mohamed-azar-portfolio