AI Systems · Retrieval · Voice Interfaces

Rayyan Habeeb builds AI you can benchmark.

Graduate Research Assistant at USF's Center for Urban Transportation Research, building voice-first RAG systems where every stage of the pipeline is traced, measured, and made faster.

Tampa, FL — M.S. Computer Science, University of South Florida

GitHub ↗ LinkedIn ↗
Portrait of Rayyan Habeeb
Rayyan Habeeb
AI Systems · GRA @ CUTR, USF
Published researcher · Exhibition winner · SnowPro certified
01 · Experience

Where the numbers come from

Jan 2026 — Present

Graduate Research Assistant

CUTR · University of South Florida
  • Transportation-safety queries at 1.16s total latency — async FastAPI parallelizing Deepgram STT/TTS with Gemini 2.5 Flash inference.
  • <400ms retrieval over 21,000+ char contexts via hybrid RAG: Qdrant vectors + BM25, fused with RRF.
  • p95 generation cut to 779ms (−35%) with granular tracing and a concurrent multi-model eval endpoint (tokens/sec, cost-per-request).
Sep 2025 — Dec 2025

Student Assistant

CUTR · University of South Florida
  • Manual review time cut 80% at 91.7% accuracy across 150GB of footage — automated pedestrian detection with OpenCV + PyTorch.
  • Processing latency down 40% via a distributed video-analytics pipeline on Lambda triggers + Fargate clusters.
  • Research timelines accelerated 86% by automating pedestrian–vehicle interaction analysis with Pandas.
Mar 2023 — Apr 2024

Full Stack Developer

Creative Technosoft Systems · India
  • CRUD REST APIs in ASP.NET Core (C#, EF Core, SQL Server) for internal business and reporting modules.
  • Responsive Razor views and admin pages (Bootstrap, jQuery) across 2 client applications.
  • xUnit tests + Serilog structured logging, helping QA reproduce and triage production bugs faster.
02 · Research & Projects

Systems, shipped

Three-part railroad training platform. VRAI-RAG: hybrid Qdrant + BM25 backend (RRF), GPU-accelerated OCR for manuals, acronym expansion, fallback chain-of-thought. ChatUI: web UI with interactive source citations, live backend metrics, side-by-side model comparison. XR Assistant: hands-free in-simulation voice guidance in Unity (Sentis, ONNX).

QdrantBM25 + RRFFastAPIUnity SentisONNX

Roman Urdu ↔ English corpus: 75,146 sentence pairs, 1.4M+ tokens, built with a hybrid synthetic pipeline (GPT-3.5, Claude Opus). MT baselines at 39.9 BLEU via fine-tuned T5-Small and mBART; human-in-the-loop eval for code-switched data.

NLPT5 / mBARTLow-resource MT

F1 Penalty Data ETL Pipeline

Data Engineering

Scrapes and structures 3,500+ FIA penalty documents — Selenium for acquisition, LangChain over Gemini and Llama to normalize regulatory text into a queryable dataset.

SeleniumLangChainETL

Verdure

Exhibition Winner

Flutter app combining ML, computer vision, and AR for smart gardening and plant-health analysis. Point the camera at a plant; get a diagnosis.

FlutterCVAR
Rayyan and team receiving certificates at the Innovatia project exhibition
Innovatia Project Expo — recognized at the exhibition, certificates in hand.

Evaluating and improving ML fairness with demographic parity and balance metrics — because accuracy isn't the only number that matters.

Fairnessscikit-learn
03 · Toolkit & Credentials

Stack

AI / Retrieval
RAGQdrantBM25LangChainOpenAI APIGeminiLlama 3.2OllamaDeepgram
Languages
PythonC#CJavaScriptSQL
Cloud / Data
AWS LambdaFargateS3EC2DockerPostgreSQLSQL ServerMongoDB
Libraries / Tools
PyTorchOpenCVFastAPIPandasNumPyxUnitSerilogGitLinux
Web
ASP.NET CoreEntity FrameworkReact.jsRazorREST

M.S. Computer Science

University of South Florida · Tampa, FL
Jan 2025 — May 2026 · GPA 3.90

SnowPro Associate

Platform Certification · Snowflake
Certified

ERUPD — arXiv:2412.17562

Published research · Bilingual parallel dataset for Roman Urdu
04 · Contact

Let's build something measurable.

Open to AI engineering, ML, and data engineering roles. If you want someone who ships pipelines and then makes them faster, get in touch.

Rayyan and team receiving certificates at the Innovatia project exhibition