-
-
Sandshell
-
Output 1
-
Output 2
-
Output 3
Inspiration
Students juggle classes, assignments, and job applications across scattered notes and apps. I wanted one assistant that holds that context and helps me plan, without handing my files to a cloud service. The Personal AI track asked for exactly that: a private assistant with memory and skills that I control.
What it does
Study & Career Copilot is an always-on personal AI that runs in a sandbox on my own machine.
- Persistent memory: facts I tell it are saved to files and survive new sessions.
- inbox-digest: I drop a document (like a syllabus) in an inbox folder. It saves a digest, extracts dated action items, and asks before adding them to my tracker.
- deadline-tracker and morning-brief: a plain-text deadline file and a short daily plan: overdue, due soon, this week, and one focus item.
- Always-on daily brief: a scheduled job writes my brief every morning at 7:00 with Nemotron Nano, without me opening the chat.
- study-planner: "Plan my week" turns my deadlines into a day-by-day plan with time estimates, using Nemotron Ultra.
- job-tracker: I paste a job posting; it saves company, role, apply-by date, and requirements, adds the date to my tracker, and tracks status (applied, interview, offer...).
- flashcards: turns my notes into flashcards and quizzes me, saving my score after every answer.
- Privacy report and forget: shows everything it remembers, every data file, the copies its background memory makes, and where my data can go. "Forget " previews every match, deletes only after I say yes, and supports undo.
How I built it
- Models on Nebius Token Factory, matched to each task through its OpenAI-compatible endpoint:
- Nemotron 3 Super for chat and skills
- Nemotron 3 Nano for the daily brief, which is simple formatting that runs every day, so it should be fast and cheap
- Nemotron 3 Ultra for weekly planning, which needs multi-step reasoning across several deadlines
A small helper script picks the model for each task and falls back to Super if a call fails.
- Runtime: OpenClaw inside an NVIDIA OpenShell sandbox, set up with NVIDIA NemoClaw. OpenClaw's built-in scheduler runs the daily brief.
- Skills: each skill is a
SKILL.mdfile. Memory and data are plain Markdown files I can open and inspect. - Privacy: OpenShell limits which folders the agent can write to and which hosts it can reach. Inside the sandbox the API key is only a placeholder that the gateway swaps for the real key, so my key never enters the sandbox. Rules like "ask before deleting" are instructions to the model, not enforcement, and the README says so.
- Nebius Serverless Jobs: the brief and planner are also packaged as a container with a guide for running them as a Nebius Serverless Job.
Challenges
- Setup on an 8 GB Mac: the OpenShell gateway kept crashing because it looks for
/var/run/docker.sock, which Colima doesn't create. A symlink fixed it. - Making the model follow skills: formats drifted, and the model sometimes answered from its own memory instead of using my skill. Putting an output template first, using distinctive trigger phrases, and adding a short routing section to the agent's instructions fixed most of it.
- Date math: the model guessed dates wrongly, so scripts compute dates and days remaining, and the model only formats or plans with them.
- Reasoning models need room: Ultra spends tokens thinking first, so my weekly plan was cut off mid-week until I raised the token budget.
- A forget that didn't ask: in testing, "forget X" deleted lines without confirmation. I made it two-step (preview, then delete only after a yes) and added undo.
- Dropped connections: after running for a day, the agent's model calls started failing mid-reply even though Token Factory answered direct requests fine. Restarting the agent gateway fixed it.
What I learned
Small, explicit skills with templates beat long instructions. Anything that has to be exact (dates, scores, deletions) belongs in a script, not the model. Picking the right model size per task saves time and credits. And sandbox policy gives real guarantees, while prompt rules only guide behavior.
What's next
Telegram delivery for the morning brief (already wired, needs a bot), PDF support in the inbox, syncing the Serverless Job back to the sandbox, and a hosted demo.
Limitations
It runs locally in a sandbox, so there is no hosted multi-user demo. The code, setup steps, and demo video are linked. NemoClaw and OpenClaw are alpha software and change quickly.
