Developer tools, AI infrastructure, and applied machine learning — from open-source CLIs and MCP servers to computer-vision research and products with real users.
21 projects
CLI + TUI for Bare-Metal AI Infrastructure
See what your bare-metal machine is doing — right now.
Built a read-only Go CLI + TUI that inspects a single AI box — host, GPUs, Docker, and local LLM runtimes (Ollama, vLLM, SGLang, Triton) — and correlates them into one snapshot, rendered as CLI tables, a live terminal dashboard, JSON, or six MCP tools for coding agents. Cross-platform, packaged via Homebrew, APT, and a no-sudo install script.
Beautiful, fully configurable bounding boxes for object detection — one Style object, 22 presets, a live Studio, and one call to render.
The open-source canvas for document AI — drag open OCR models onto a canvas, wire them together, and turn any PDF into validated JSON on your own hardware.
An MCP server that hands an AI agent a styled SVG map of Jordan — 12 governorates, 122 hospitals, five typed tools, no API key and no network at call time.
A quiet, local command center for people juggling more than one project at a time — what's open, what belongs where, what must happen today, and what's already done.
Find the smallest realistic change that breaks your computer-vision model — local-first robustness testing from failure-boundary search to full-stream video evaluation.
Prometheus metrics for Open WebUI — polls the REST API, no database access required.
Type a prompt, watch an AI think — live attention, logit-lens predictions, and token probabilities, streamed token by token from a real forward pass.
Prometheus metrics for self-hosted Qdrant — the per-collection, per-shard, and cluster data the native /metrics endpoint never exposes.
A drag-and-drop canvas for building ML pipelines without code — rebuilt from a single-file Streamlit app into a multi-service SaaS.
A Python package that turns a whole project's files into one well-structured LLM prompt.
Upload a pair of retinal photos, get a graded diabetic retinopathy report back — formerly RetinAI.
Upload a prescription, get it explained back — a fine-tuned T5 model that answers questions about your own medication.
One function call to fix Arabic text alignment across Streamlit components.
Real-time face detection, emotion recognition, and multi-object tracking from a webcam feed — RetinaFace + a residual masking network + SORT.
TF-IDF search over ~5,000 Arabic articles, built with a custom preprocessing pipeline and a hand-rolled tokenizer.
A YOLOv5 card detector feeding a from-scratch Tarneeb round-winner engine — built and debugged with a teammate over one semester.
The Streamlit app built on top of Face Blurring — upload an image, pick YOLOv8 or Haar Cascade, get the faces blurred in-browser.
Two models, head to head — a fine-tuned ResNet50 and a custom CNN, trained on the same 12K-image dataset, evaluated the same way.
Haar Cascade vs. a YOLOv8 model trained from scratch on a custom Roboflow dataset — the research project behind Blurify.AI.
Restyle a photo as Van Gogh, Kandinsky, or three other artists in one click, powered by Google Magenta's pretrained arbitrary-style-transfer model.
