The invisible
engine.

When an external API your product depends on starts failing or degrading, you usually find out from user complaints, not your own monitoring — and by the time you notice, you've already lost valuable diagnostic time.

An asyncio worker checks the status of every monitored service in parallel every 30 seconds and reports results to the backend over HTTP — it never touches the database directly. The backend automatically detects status transitions (healthy → degraded → down, and back), opens and closes incident records with their exact duration, and broadcasts every change live to all connected browsers via WebSocket, no refresh needed. Latency charts use adaptive downsampling instead of fixed hourly buckets, so a newly added service never shows an empty chart during its first hour.

Cloud-based AI assistants mean sending every conversation to third-party servers, paying per token, and depending on the service staying available — with no real control over the model, the history, or the privacy of what's shared.

Trenchcoat runs entirely on the local machine: a local LLM via Ollama for text and code, and Stable Diffusion (Forge) for images, automatically orchestrated based on each message's intent, without the user having to switch modes manually. An intent classifier decides whether a message is text, a new image, or an edit of the previous image — with a fallback model for cases where pattern matching isn't enough. Image edits have an intensity control (subtle / normal / aggressive) to balance fidelity against how much the result is allowed to change.

Manually reviewing dozens of CVs against a job opening is slow, inconsistent across reviewers, and leaves no clear trail of why someone was rejected.

An LLM analyzes each CV against the job description and returns a score, a requirement-coverage percentage, and a breakdown of matched and missing skills — with consistency rules enforced via prompt so the score and the displayed reasons always tell the same story. It also extracts email, phone, and contact links directly from the PDF via regex.

Generic chatbots burn through tokens answering without relevant context.

RAG pipeline with ChromaDB: only relevant context is sent to the AI, reducing token usage. Redis cache for repeated queries.