---
title: Plane Mode Scholar
emoji: ✈️
colorFrom: blue
colorTo: indigo
sdk: gradio
app_file: app.py
pinned: false
license: mit
short_description: Fine-tuned Nemotron 4B study coach + MemoryAgent
suggested_hardware: l4x1
startup_duration_timeout: 1h
models:
  - GuusBouwensNL/plane-mode-nemotron-4b-study-coach
  - nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF
preload_from_hub:
  - nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf
  - GuusBouwensNL/plane-mode-study-coach-lora-gguf plane-mode-study-coach-lora.gguf
tags:
  - Backyard AI
  - Off-Brand
  - Sharing is Caring
  - Field Notes
  - Best Agent
  - Off the Grid
  - Well-Tuned
  - track:backyard
  - sponsor:nvidia
  - achievement:offgrid
  - achievement:welltuned
  - achievement:offbrand
  - achievement:llama
  - achievement:sharing
  - achievement:fieldnotes
---

# Plane Mode Scholar

Build Small Hackathon entry — Backyard AI track, Nemotron Quest.

I built this for a grad student who studies on planes and in libraries where Wi-Fi drops constantly. Most study apps forget everything between sessions. This one remembers what you got wrong, what exam is coming up, and how you like things explained.

Hit **FLY** once. The agent packs your materials, pulls up due reviews, explains a topic with citations from your notes, and runs a quiz. No tab-hopping.

The coach is a fine-tuned [Nemotron 3 Nano 4B](https://huggingface.co/GuusBouwensNL/plane-mode-nemotron-4b-study-coach) — SFT on real study conversations. UI is custom HTML on `gr.Server`, loosely based on my [SwarmGrid](https://github.com/GJB99/SwarmGrid) project.

## Demo

![Plane Mode Scholar autopilot demo](https://github.com/GJB99/plane-mode-scholar/raw/demo-v1/docs/demo/plane-mode-scholar-demo.gif)

One tap FLY, then watch it plan → retrieve → explain → quiz. Runs offline with llama.cpp if you need that.

| | |
|---|---|
| Space | https://huggingface.co/spaces/build-small-hackathon/plane-mode-scholar |
| Demo MP4 | https://github.com/GJB99/plane-mode-scholar/raw/demo-v1/docs/demo/plane-mode-scholar-demo.mp4 |
| Social post | https://x.com/GuusBouwens/status/2066670913467400284 |
| Field notes | [docs/field-notes.md](docs/field-notes.md) |
| Badge writeup | [docs/badges.md](docs/badges.md) |
| Trace dataset | [docs/agent-traces-dataset.jsonl](docs/agent-traces-dataset.jsonl) |

## Badges

| Badge | What we did |
|-------|-------------|
| Well-Tuned | LoRA SFT on Nemotron 4B → [plane-mode-nemotron-4b-study-coach](https://huggingface.co/GuusBouwensNL/plane-mode-nemotron-4b-study-coach) |
| Off the Grid | Local inference via llama.cpp GGUF — no cloud LLM calls |
| Off-Brand | Custom dashboard, not default Gradio tabs |
| Llama Champion | `./scripts/start_llamacpp.sh` runs the fine-tune through llama-server |
| Sharing is Caring | Open agent traces + `/export_trace` endpoint |
| Field Notes | [docs/field-notes.md](docs/field-notes.md) — what broke, what surprised me |
| Best Agent | `StudyAgent` does monitor → plan → act without user clicking through steps |
| Nemotron Quest | Nemotron 3 Nano family throughout (4B coach, 30B MoE fallback) |

Fine-tune details: [docs/finetune.md](docs/finetune.md). Colab notebook in `notebooks/`.

## Try it locally

```bash
pip install -r requirements.txt
export HF_TOKEN=your_token
python app.py
```

Open `http://localhost:7860`.

Fully offline path — download the GGUF, start llama-server with the LoRA, point the app at it:

```bash
./scripts/setup_llamacpp_stack.sh   # first time only
./scripts/start_llamacpp.sh           # terminal 1
PMS_INFERENCE_BACKEND=llamacpp python app.py   # terminal 2
```

## Models

| Role | Model |
|------|-------|
| Coach (default) | `GuusBouwensNL/plane-mode-nemotron-4b-study-coach` |
| Base | `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16` |
| Heavy fallback | `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8` |
| Offline GGUF | `nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF` |
| Embeddings | `sentence-transformers/all-MiniLM-L6-v2` |

Repo: [github.com/GJB99/plane-mode-scholar](https://github.com/GJB99/plane-mode-scholar)

MIT