Skip to content
Claude CodegreenfieldBackend40 files · +1936 −736

Plan and scaffold TutorDesk AI for a Hugging Face hackathon

Session from tutordesk-agent-traces on Hugging Face by naazimsnh02 · MIT · condensed by Coders Talk

49msession
0interventions
GOAL

Turn a rough PRD into a hackathon-optimized concept that maximizes sponsor/award coverage for the Build Small Hackathon, then scaffold and start building the actual TutorDesk AI app (Gradio + Modal-hosted models).

  1. Prompt · +0
    Check PRD.md and hackathon requirements/sponsors/awards, and figure out which 5 features maximize winning potential, ignoring OpenAI/NVIDIA sponsors.
  2. Agent did · +0m
    Researched hackathon page and sponsors, proposed 'sponsor stacking' strategy, asked clarifying questions (JetBrains, model size), then rewrote PRD.md v2.0 around 5 features each targeting a different sponsor/award.
  3. Prompt · +7m
    Asked whether MiniCPM-V 4.6 (1B) would hurt quality vs 4.5 (8B), and questioned the purpose of Aya/Tiny Aya.
  4. Prompt · +14m
    Asked if Feature 5 (Parent Progress Report) was actually useful in HF Spaces or should be replaced; chose Photo Auto-Grading with fine-tuned Qwen instead.
  5. Agent did · +17m
    Redesigned Feature 5 as a two-model pipeline (MiniCPM-V vision extraction + fine-tuned Qwen3-4B grading with Indian marking conventions) and updated PRD sections, fine-tune scope, roadmap, and checklist accordingly.
  6. Prompt · +20m
    Asked to check for existing datasets to reuse for fine-tuning, and to consider narrowing scope to fewer subjects/classes.
  7. Agent did · +20m
    Searched Hugging Face for existing Indian curriculum datasets, found NCERT_Dataset (124k rows, MIT license) as a workhorse base, then created CLAUDE.md, implementation_plan.md, and progress.md to track the build.
  8. Break · 35m away, not counted
  9. Agent did · +23m
    Scaffolded the full repo structure: agents, models, pipelines, utils, traces, data prep, finetune configs, Modal serving stub, and app.py, verifying everything byte-compiles.
  10. Prompt · +32m
    Explained laptop has no memory for local inference; asked whether Cohere/FLUX should be self-hosted on Modal instead of using APIs, and to check model choices for size limits.
  11. Agent did · +39m
    Locked architecture on all-Modal self-hosting (Tiny Aya 3.35B, MiniCPM-V, Qwen3-4B, FLUX ~27B total), wired serving/modal_app.py and model clients, then wrote real Phase 1 code (Modal functions, PDF export, weekly pack pipeline, Gradio app) and confirmed it compiles.
  12. Fail · +49m
    Deploying to Modal failed with a KeyError: 'str' from modal.parameter() not supporting string types.
  13. Outcome · +49m
    Fixed by removing modal.parameter() in favor of module-level config constants; serving/modal_app.py imports cleanly and is ready for redeploy, with full repo scaffolded across agents/models/pipelines/serving.
VERDICT

I'd start by pasting the hackathon page and PRD together and asking directly for a sponsor-coverage table, since that's what actually shaped every later decision. Pushing back on Feature 5 twice (quality of MiniCPM-V 4.6 vs 4.5, then usefulness of the manual Parent Report) was the correction that mattered most — it turned a weak filler feature into the two-model auto-grading pipeline that became central. By the end we had a fully scaffolded repo and a Modal deploy bug caught and fixed in the same session, but no working end-to-end run yet.

I tried this · 0

Discussion · 0

How Builds work