asanwari/secret-student

<div align="center">

Secret Student

A Pokemon-inspired AI school game built for a niece who loves games and hates school

    

Learn at school. Get briefed by "Grandma". Defeat a villain with what you learned.

<p align="center"> <img src="frontend/static/assets/world-map-agent.png" alt="Pixel-art Secret Student main map with school, home, and headquarters"> </p>

<table> <tr> <td width="50%"><img src="frontend/static/assets/classroom-blank.png" alt="Pixel-art Secret Student classroom"></td> <td width="50%"><img src="frontend/static/assets/bedroom-agent.png" alt="Pixel-art Secret Student bedroom"></td> </tr> <tr> <td align="center"><strong>School: lessons, teacher chat, and quizzes</strong></td> <td align="center"><strong>Home: review and mission briefings</strong></td> </tr> </table>

</div>

Secret Student turns an AI-generated lesson into a retro 2D mission. The player is a student by day and a secret agent after class. Lessons become intelligence, quizzes become training, and the final assessment becomes a boss battle.

The whole AI stack can run locally through two small GGUF models and llama.cpp. There is no required cloud model API, and the custom Phaser interface looks and plays like a game rather than a default chatbot.

Hackathon Submission

Secret Student is submitted to both Build Small tracks. The required organization submission is [build-small-hackathon/secret-student](https://huggingface.co/spaces/build-small-hackathon/secret-student).

[!IMPORTANT] The organization Space cannot currently provision GPU hardware, so the live submission runs the game in a CPU-hosted Space and serves the current text and vision models through a routed Modal GPU deployment. This is a hosting workaround, not a model-size requirement: the same source can run both models in one local llama.cpp deployment when GPU hardware is available.

Team

• —asanwari

Judges: Inference and GPU Setup

[!CAUTION] Please read this before evaluating latency. The current submission calls a Modal-hosted llama.cpp deployment because the organization Space cannot provision a GPU. Modal scales down when idle, so the first lesson or handwritten-answer check after an idle period can be significantly slower while the GPU starts and the GGUF models load. This is cold-start latency, not the normal warm inference path. If you encounter LLM unavailable errors, please retry.

The current model pair uses about 10 GB of VRAM when loaded together. It can run on one GPU with at least 16 GB of VRAM, including:

• —NVIDIA GeForce RTX 4060 Ti 16 GB
• —NVIDIA GeForce RTX 4070 Ti SUPER 16 GB
• —NVIDIA GeForce RTX 4080 / 4080 SUPER 16 GB
• —NVIDIA GeForce RTX 3090 or RTX 4090 24 GB
• —AMD Radeon RX 7800 XT or 7900 GRE 16 GB
• —AMD Radeon RX 7900 XT 20 GB or 7900 XTX 24 GB

LLM Runtime Options

LLM_RUNTIME selects the inference topology:

mock                     deterministic local content with no model server
external                 OpenAI-compatible external routes, including Modal
embedded_llamacpp        one llama.cpp server beside the app
embedded_dual_llamacpp   separate local text and vision llama.cpp servers

The submission currently uses external. The fully local GPU path is embedded_dual_llamacpp.

How to Test Completely Locally

The Docker image already includes CUDA-enabled llama.cpp, and app.runtime already knows how to launch and health-check separate text and vision servers. No code changes or additional inference service are required.

1. 1.Run on a machine, Docker host, or Space with a suitable GPU.
2. 2.In .env or Settings -> Variables and secrets, replace the external runtime configuration with the values below.
3. 3.Restart the app. On the first boot, allow time for the model downloads unless the GGUF files have already been placed in persistent /data/models storage.

Variables:

LLM_RUNTIME=embedded_dual_llamacpp
LLM_PROVIDER=openai_compatible

LLM_MODEL=nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF:Q4_K_M
LLAMA_CPP_MODEL_REF=nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF:Q4_K_M
LLAMA_CPP_CTX_SIZE=8192
LLAMA_CPP_GPU_LAYERS=999
LLAMA_CPP_THREADS=8
LLAMA_CPP_PARALLEL=1
LLAMA_CPP_STARTUP_TIMEOUT=900

VISION_LLM_MODEL=openbmb/MiniCPM-V-4_5-gguf:Q4_K_M
VISION_LLAMA_CPP_MODEL_REF=openbmb/MiniCPM-V-4_5-gguf:Q4_K_M
VISION_LLAMA_CPP_CTX_SIZE=4096
VISION_LLAMA_CPP_GPU_LAYERS=999
VISION_LLAMA_CPP_PORT=8002

If the runtime does not have preloaded model files, delete LLAMA_CPP_MODEL_PATH, VISION_LLAMA_CPP_MODEL_PATH, and VISION_LLAMA_CPP_MMPROJ_PATH. Their absence tells llama.cpp to download the configured GGUF repositories. For faster restarts with persistent storage, preload the files and set:

LLAMA_CPP_MODEL_PATH=/data/models/nemotron-3-nano-4b/NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf
VISION_LLAMA_CPP_MODEL_PATH=/data/models/minicpm-v-4_5/MiniCPM-V-4_5-Q4_K_M.gguf
VISION_LLAMA_CPP_MMPROJ_PATH=/data/models/minicpm-v-4_5/mmproj-model-f16.gguf

LLM_BASE_URL, VISION_LLM_BASE_URL, LLM_API_KEY, and VISION_LLM_API_KEY belong to the current Modal configuration. They are ignored by embedded_dual_llamacpp and may be removed when switching to fully local inference.

Configurability Built Into the Project

The repository includes set_env_for_space.py and set_env_local.py to apply external Modal routing in one command. Full model-preloading, Modal, Docker Compose, and trace instructions are in the Technical Guide.

The application can separate the CPU-hosted game from inference. In external mode, independent OpenAI-compatible routes are assigned through LLM_BASE_URL and VISION_LLM_BASE_URL. The included YAML-driven Modal deployment runs both models, plus optional additional models, in one GPU container and exposes each at its configured subroute.

Submission Links

The achievement:welltuned tag is intentionally not claimed: the current models are configurable base models, not a fine-tuned model published by this project.

The Game

1. 1.Create your agent. Choose a student name, secret codename, grade, and the hair, shirt, and pants colors for the backpack-wearing character.
2. 2.Go to school. Pick a topic and receive a structured, age-appropriate lesson generated for that student.
3. 3.Ask the teacher. Follow-up chat stays grounded in the current lesson.
4. 4.Take the quiz. Type an answer or draw it in the notebook. Feedback remains visible until the player chooses to continue.
5. 5.Return home. Review material at the desk, then answer the secret phone for a comic-style briefing from the handler known only as Grandma.
6. 6.Enter headquarters. Fight a villain by answering harder questions based on the lesson. Correct answers damage the boss; mistakes cost health.
7. 7.Start another mission. Return to the map and learn something new.

Small Models, Separate Jobs

flowchart LR
    P[Player] --> UI[Custom Phaser game]
    UI --> API[FastAPI game API]
    API --> DB[(SQLite progress)]
    API -->|Lessons, chat, typed grading| TEXT[NVIDIA Nemotron 3 Nano 4B GGUF]
    API -->|Handwriting only| VISION[MiniCPM-V 4.5 GGUF]
    TEXT --> L1[llama.cpp text server]
    VISION --> L2[llama.cpp vision server]
    L1 --> TRACE[Validated traces]
    L2 --> TRACE

One approximately 4B NVIDIA Nemotron text model handles lesson generation, teacher chat, question creation, and typed-answer grading. A separate OpenBMB MiniCPM-V model is called only when pixels need to become an answer. Both are quantized and served by llama.cpp through OpenAI-compatible local endpoints.

This split keeps the common path fast and leaves vision work to a model designed for it. Model repositories, files, ports, context sizes, and GPU layers are all environment-configurable; Nemotron and MiniCPM-V are defaults, not hard dependencies.

The default pair uses about 10 GB of VRAM when loaded together and is comfortable on a 16 GB GPU with conservative context settings. Practical consumer options include the NVIDIA RTX 4060 Ti 16 GB, RTX 4070 Ti SUPER 16 GB, RTX 4080 / 4080 SUPER 16 GB, RTX 3090 or RTX 4090 24 GB, AMD Radeon RX 7800 XT or 7900 GRE 16 GB, and AMD Radeon RX 7900 XT 20 GB or 7900 XTX 24 GB. Partial CPU offload can reduce VRAM requirements at the cost of latency.

Reliable Generated Content

The model never writes directly into the game. Secret Student treats every completion as untrusted input:

flowchart LR
    A[Prompt + JSON Schema] --> B[Model completion]
    B --> C[Extract and parse JSON]
    C --> D[Normalize safe aliases]
    D --> E[Pydantic + content rules]
    E -->|Valid| F[Persist lesson]
    E -->|Invalid| G[Repair prompt with exact errors]
    G --> H[Validate again]
    H -->|Valid| F
    H -->|Invalid| I[Fail with trace ID]

The validation layer enforces six to ten progressive lesson steps, a 150-word limit per step, concrete expected answers, exact question counts, and no references to unavailable maps, images, videos, or worksheets. Common JSON formatting damage is repaired locally. Deeper schema or content failures are sent back to the model with the exact validation errors for one constrained self-correction pass. Only a valid result is saved.

Every model call can produce a durable trace containing timing, parsing, normalization, validation, repair, and final status. Secrets are redacted and image payloads are replaced with hashes.

Sponsor Stack

Achievement Evidence

Built for the Judging Criteria

• —Technical implementation: two specialized local models, explicit routing, validated structured generation, model-assisted repair, durable traces, persistence, and responsive desktop/tablet controls.
• —Model choice and use: NVIDIA Nemotron 3 Nano 4B handles language-heavy work while MiniCPM-V is invoked only for handwriting. Game rules remain deterministic application code rather than being delegated to the models.
• —Creativity: the lesson, quiz, home briefing, and boss fight form one coherent Pokemon-inspired adventure instead of a chatbot wrapped in educational copy.

Run Locally

The fastest local preview uses deterministic mock content:

uv sync
cp .env.example .env
LLM_RUNTIME=mock uv run uvicorn main:app --reload --host 0.0.0.0 --port 7860

Open http://127.0.0.1:7860.

For the full dual-model setup, Hugging Face Space configuration, Modal deployment, model preloading, tracing, API routes, and tests, see [Technical Guide](docs/TECHNICAL.md).

Project Links

• —GitHub source and Codex-assisted history
• —Technical guide
• —Medium field notes
• —Public LLM trace dataset
• —NVIDIA Nemotron 3 Nano 4B GGUF
• —OpenBMB MiniCPM-V 4.5 GGUF
• —llama.cpp