Meet your chief of staff.
It tells you what you are quietly letting drop, measured against what your company says your job actually is. Built first for a new engineering manager, it works for anyone whose role comes with written expectations: a product manager, a sales lead, a new director, a founder, a team of one.
The problem. No task manager can solve this. Task managers list what you decided to write down. They have no idea what you never wrote down at all. That blind spot is where careers stall. What this replaces is the pattern recognition of a great manager, available during your first six months, when nobody is doing it for you.
How it works. After each meeting you type or speak a short memo. A model pulls out who owes what to whom and by when, and nothing is saved until you approve it. Plain database queries then compare your record against your handbook and report which parts of your job your notes have gone quiet about. Never how well you did them. Only whether they came up at all. Of the 39 expectations drawn from the handbook, 27 are marked checkable. The other 12 need human judgement, so the app refuses to fake one.
Every morning, a ranked list. Four signals decide what comes next: the deadline, how much your role says it matters, the size of the job, and whether the area it serves is going quiet. The weights are instructions to a model, not coefficients. Nothing is multiplied out and nothing carries a score, because an item labelled 0.62 gives a reader nothing to argue with.
Under the hood. LangChain and LangGraph drive an agent with seven tools. ChromaDB with a multilingual embedding model runs on your own machine, so an English question finds a German handbook page and retrieval costs nothing. Streamlit gives you eight screens. Notion supplies the handbook. A Telegram bot captures written or spoken notes while you move, transcribed locally by Whisper instead of a cloud service, because a cloud service would hear the audio before any guard had read the words.
Tested by trying to make it flatter me. Three memo sets: clean, with problems planted, and the same problems rewritten in cheerful language. The right answer is known before anything runs. It finds all four planted signals in both sets, invents nothing on the clean set, and the cheerful wording changes the outcome not at all. Better still: a lazy strategy that reads no memo and simply claims everything has drifted earns the same headline number, at a precision of 0.035. Proving that a metric can be gamed is stronger evidence than passing it.
2,588 tests, six files of which exist only to forbid things: no rating, no total, no blended score, no judgement about a colleague, and no private note reaching anything that sends.
Ethics came first and shaped the build. German law makes recording another person's conversation a criminal offence, so there is no meeting recorder. You dictate your own summary, alone. Private notes about working relationships sit behind a wall in the code that modules facing outward have no route through. Notes that stray onto a colleague's health are refused outright, and nothing is kept. No real company and no real person appears anywhere in the repository.
