Obscure catches secrets and personal data in your code and text before it ever reaches a cloud LLM like Claude, ChatGPT, or Gemini. Regex rules catch known patterns; Mahfuj-0.1, a small on-device AI model, catches the rest — names, addresses, anything regex can't reliably see. Everything runs locally. Nothing you scan ever leaves your machine.
Two layers, working together, both running entirely on your own hardware.
Layer one — regex. Fast, deterministic pattern matching catches AWS keys, GitHub tokens, API keys, JWTs, private keys, passwords, and other structured secrets. Also flags any high-entropy string that looks like a secret, even if it doesn't match a known pattern.
Layer two — Mahfuj-0.1. A fine-tuned 0.5B parameter model (built on Qwen2.5) trained specifically to catch what regex can't: personal names and street addresses buried in code comments, config files, logs, and prose. It runs as a highlight-only fallback — findings are flagged for your review, never silently auto-redacted, since a missed secret is far worse than a false alarm.
Two commands cover the core workflow: scan to see what Obscure finds, and redact to produce a redacted copy plus a private map file to reverse it later.
A second NanoBots model, Mahfuj Cmp 0.1 — an AI intent-detection model for compilation — is currently in development.
Free and open-source
No account, no upload, no cloud step. Code and model, both downloadable.
Download the code -> ModelThe fine-tuned detection model behind Obscure's AI layer, hosted on Hugging Face.
View on Hugging Face -> Live walkthroughGet in touch for a short on-site demo, or ask about integrating Obscure into your workflow.
Request a demo ->Live demo — connect your own API
Fleet Watch — a live simulated delivery run
A heavier cloud model plans the route and issues the mission. Mahfuj 1 Nano rides on the drone itself, watching three numbers — battery, distance to the nearest obstacle or drone, and altitude deviation from the planned path — and reacts in milliseconds if any of them cross into unsafe territory. Below, a delivery run generates synthetic telemetry on a timer, exactly like a real flight would, and the model reacts live.