Ahmad Freijeh

Self-hostable, multi-layer semantic caching library for Node.js LLM apps. Runs two cache layers — exact Redis match then Qdrant semantic search — before falling through to your LLM, so repeated or similar queries cost nothing. Caches embeddings to avoid re-embedding, returns full latency breakdown and estimated cost saved, and works with any LLM provider. Named after déjà vu.

• TypeScript
• Node.js
• Redis
• Qdrant
• OpenAI Embeddings

You're looking at it. Bauhaus-inspired personal portfolio built with Next.js, Tailwind CSS, and Framer Motion. Retro editorial design, zero templates — designed and shipped in a single focused sprint.

• Next.js
• TypeScript
• Tailwind CSS
• Framer Motion

ML prediction SaaS platform built with Next.js and FastAPI. Features multi-model training pipelines, confusion matrix visualisation, dataset versioning, and real-time inference — backed by PostgreSQL and scikit-learn.

• Next.js
• FastAPI
• Python
• PostgreSQL
• scikit-learn

Field and emergency use AI system designed for zero-connectivity environments. Features Arabic language support, offline RAG, speech-to-text, and vector search. Runs local LLMs via Ollama — Qwen (1.8B, 7B) and DeepSeek (1.3B, 7B) — with custom parameters for context length, temperature, and memory constraints optimized for edge hardware.

• Python
• Ollama
• Qdrant
• FastAPI
• Whisper
• Node.js

Vision AI that tracks how old your banana is — trained on images of bananas at every stage of ripeness using a ResNet model, it estimates age from the moment of purchase. Point a camera, get a verdict.

• Python
• ResNet
• PyTorch
• Image Classification
• FastAPI
• Streamlit

Handwritten prescription OCR API — upload an image, get back drug name, dosage, frequency, and prescriber as structured JSON. Pipeline runs TrOCR for handwriting recognition, fuzzy-matches drug names against the FDA NDC dictionary, and uses Flan-T5 for field extraction. Includes a LoRA fine-tuning pipeline via PEFT that trains only ~0.47% of parameters.

• Python
• TrOCR
• Flan-T5
• LoRA / PEFT
• FastAPI
• OpenCV
• rapidfuzz

AI powered relationship text analyzer. Paste any message from your partner and find out if you are actually okay — or if you should start preparing yourself. Powered by BERT fine tuned on real relationship data sourced from Reddit and HuggingFace emotion datasets. Built with PyTorch, HuggingFace Transformers, FastAPI, and Streamlit to detect four danger levels: Safe, Suspicious, Danger, and It's Over.

• Python
• BERT
• PyTorch
• HuggingFace Transformers
• FastAPI
• Streamlit