Rayane Goghrod
I'm Rayane, an AI engineer. While completing my engineering degree, I worked with companies to gain real-world experience on various projects: fine-tuning speech-to-text models for specific use cases, ETL pipelines, a fraud detection pipeline for financial transactions, and a personalized AI assistant (full-stack app).
During my final year, I joined a dual degree program at Mississippi State University, where I'm currently completing a master's degree, graduating in December. During my first semester, I worked at the AI lab building a Lego education kit robot with AI features and guiding students on how to reproduce and improve it. I also participated in research projects, being second author on one paper and first author on another, both under review at conferences.
Experience
research · Starkville, MS
- Research under Prof. Zhiqian Chen. First author on one paper and co-author on another, both currently under review.
- Assisted in his intro ML course: built a demo project (Lego robot controlled via real-time hand gesture recognition using computer vision), delivered a guest lecture, and supported students through office hours.
PyTorch · ROME · MEMIT · WISE
industry · Casablanca, Morocco
- Sole AI engineer at a 6-person startup. Built and deployed a full-stack internal AI assistant (Next.js, FastAPI, WebSockets) with persistent memory and integrations across Gmail, Calendar, Jira, and GitHub, in daily use by the whole team. Designed a "skills" system: slash-command shortcuts that inject curated prompts and context at runtime, cutting repetitive re-prompting.
- Architected a real-time fraud detection pipeline for Ghirpay, a fintech product in development. Chose unsupervised learning to handle the cold start problem (no labeled fraud data), Redpanda for low-latency streaming, and Isolation Forest for anomaly detection, with a feature store on MongoDB.
Next.js · FastAPI · WebSockets · Redpanda · MongoDB · Docker · OAuth2
industry · Casablanca, Morocco
- Fine-tuned Whisper on French news audio enriched with Moroccan named entities, cutting name-spelling errors from near-total failure to
12%on held-out data. - Built a job market data pipeline: crawled and scraped listings from multiple sites, parsed and classified unstructured content into structured JSON via LLM, and stored in a database with semantic search over job offers.
Whisper · PyTorch · Hugging Face · Python · ffmpeg · RAG
Projects
Built when LLMs first got reliable at tool calling. Wanted a personal assistant across Discord and Telegram: web search, task management, weather, persistent conversation memory, and voice I/O via Whisper and ElevenLabs. Designed as a multi-agent system since a single agent couldn't handle that many tools reliably at the time.
Python · Whisper · ElevenLabs · Discord API · Telegram API
Built after scrambling to finish a presentation on a deadline. Discovered sli.dev and Reveal.js and figured AI could generate the whole thing. Full-stack generator (React + FastAPI) with retrieval-grounded web search for accuracy and automated image generation. Prompt in, slide deck out.
React · FastAPI · RAG · Python
Built 4 years ago to automate daily video compilations from top-voted subreddit posts, inspired by Daily Dose of Internet. The interesting problem: compositing mixed aspect ratios (4:3, 16:9, vertical) onto a 1080p canvas without naive stretching, solved with derived scaling math, a problem that took days to work out before LLMs made it trivial.
Python · ffmpeg · YouTube API
Open source