Inspiration
Every company runs on workflows that live only in people's heads: how incoming invoices get booked, how a support escalation is handled, how a release is cut. Nobody documents them, because documenting is slower than doing. We wanted a system that watches real work with consent, understands it, and writes the playbook, so those workflows can be reviewed, improved, and handed to an agent or a new hire.
The raw material is click-by-click logs, screenshots and meeting audio from employees' desktops. That can never be a black-box SaaS. It has to run on open models, on infrastructure the company chooses. That is exactly what NVIDIA Nemotron on Nebius Token Factory makes possible, and it is why the whole thing is now open source under MIT at github.com/braincache-me/braincache-os.
What it does
1. BrainCache for macOS is a native Windows/Mac menu-bar app and privacy-first activity recorder. It logs app switches, clicks resolved to accessibility roles and labels, keyboard shortcuts, window screenshots, and meeting or dictation transcripts. Password fields are never read, plain keystrokes are never stored (only shortcuts like ⌘C), password managers and banking apps are excluded by default, and everything stays in a folder the user picks. A read-only braincache CLI reads the same data without the app running.
2. Workflow Studio is a local Node.js + React console for team leads and operations. Drop in the daily logs, press Initialize, and a Nemotron agent reads the evidence through tool calls, groups events by business outcome ("Book incoming invoices correctly") rather than by app, builds each task's chronology, and reports effort metrics: active minutes, attempts, corrections, app switches, friction points. From any task you generate a generalized SKILL.md with {placeholders} and "where to find it" hints, ready for an agent runtime or a colleague. You can set target goals up front, edit any extracted process, and ask follow-up questions in a chat grounded in the same logs.
How we built it
Inference: NVIDIA Nemotron through Nebius Token Factory. The API is OpenAI-compatible, so one client path serves four models, routed by job:
| Job | Model |
|---|---|
| Per-event labelling, clip classification, quick follow-ups | Nemotron 3 Nano |
| Agentic analysis loop, skill authoring, in-app assistant | Nemotron 3 Super |
| Final cross-app goal assignment | Nemotron 3 Ultra |
| Voice and meeting transcription, screenshot inspection | Nemotron 3 Nano Omni |
Embeddings are Qwen3-Embedding-8B on Token Factory, truncated Matryoshka-style to the existing 256-dimension vector store.
Challenges we ran into
- Turning noisy cross-app click streams into outcome-level tasks rather than "used Excel for 40 minutes". The fix was prompting the model to ask why each action happened, connect events across apps and time gaps, and never name a goal after an app or a verb.
- Moving off the Responses API. Tool-call bookkeeping is a bit hard: every call must get exactly one matching result on the next turn or the run dies mid-analysis, including calls skipped for budget.
- Privacy on the capture side: secure-field detection, exclusion lists, fragmented writes and a heartbeat so a crash never loses a recording or a transcript.
Accomplishments that we're proud of
An end-to-end loop from raw desktop activity to a reviewed, generalized skill file, with effort metrics a manager can act on, running entirely on open models. The benchmarks that we used demonstrated significant improvement of process automation using Braincache.
What we learned
Open models on Nebius are strong enough for evidence-driven agentic analysis, and multi-model routing is what keeps a log-heavy workload both responsive and affordable. An omni model that takes audio on the same endpoint as text removes a whole category of integration. And for an activity recorder, "your data never leaves infrastructure you control" is not a nice-to-have, it is the entire sales conversation.
Log in or sign up for Devpost to join the conversation.