When the promptstopsworking

Three engagement formats

Audit & Spec

Map your manual processes → architectural spec for an agent system. Includes AI/Tool/Code node decomposition and ROI assessment.

Build & Ship

Pipeline implementation on n8n / Claude Code / custom backend. CI/CD, observability via Langfuse, LLM fallback routes, alerting.

Embed Engineer

Part-time embedded agent engineer with your team. Help with design, reviews, fighting production fires.

Practice OS — for private practice

A format for solo practitioners (therapists, coaches, healers): a diagnosis of how you work with AI today — borrowed tools, personal practice, or sovereign practice — and the assembly of your Practice OS: the memory, knowledge base, helpers and rituals of your work. 1:1, in person in Nashville (The Healing Society) or remote.

A benchmark for people who code with AI: it reads local session transcripts and measures how the work actually went. The default run makes zero network calls; before anything is submitted it prints the exact JSON and asks. Published on npm, one command to run.

Sixteen independent interpretation engines under one umbrella: deterministic calculation first, LLM explanation on top. The governance rule is strict — with no retrieved sources the model is never called, so the system cannot invent doctrine. Shared RAG layer, single gateway to the models.

Bilingual dictation that runs on your own hardware: Russian and English, including both mixed inside one sentence. Recognition happens on-device and types straight into whatever app is in front of you — no cloud, no account, no telemetry. Windows, macOS, Linux.

A monorepo built around a source corpus: conversion into clean Markdown, a typed SDK for validating and querying it, and a desktop tutor that runs practices and reads lessons over that corpus. Works offline; the architecture is decoupled from any single tradition.