Easy Street

• 

  Actual hosted workspace: normal document upload and the original sample workflow.

• 

  Actual saved Nemotron course: exact source quotations and paragraph locations for review.

Inspiration

A good chapter can still take hours to turn into useful lessons and practice. I wanted an instructor or independent learner to keep the source close while shaping a course, checking a claim, and correcting the draft.

What it does

Easy Street turns a readable PDF, DOCX, or TXT chapter into an editable mini-course. Review the extracted passages, choose modules, assignments and quizzes, then approve the outline and selected source coverage. A separate one-click workflow uses the included original, MIT-licensed garden chapter.

NVIDIA Nemotron on Nebius Token Factory drafts the course and performs a separate support review. Evidence views show exact quotations and source locations. Missing evidence and conflicting support can remain flagged. The workflow does not search the web or retrieve outside educational material. Exact quotations and model review help inspection; they do not guarantee every claim is supported.

Edit a section, inspect its version history, or regenerate it without replacing unrelated edits. Export learner HTML, a separate answer key, or a JSON course backup.

How I built it

I extended the existing Easy Street repository. React and TypeScript provide the editor; FastAPI and SQLite handle extraction, course versions, background jobs, caching and usage accounting. The backend Nebius adapter uses nvidia/nemotron-3-super-120b-a12b for real generation and support-review calls. OpenAI support and normal bring-your-own-key local installations remain available.

The separate Fly judging deployment keeps its operator-funded Nebius key in runtime secrets. Visitors receive isolated temporary workspaces, deletion controls and 24-hour active-file cleanup. Backups may persist for their disclosed retention period. Document text, including embedded malicious instructions, is treated as untrusted source data.

Upload, token, course and concurrency limits bound generation. Atomic usage reservations, bounded retries, cached reuse and an emergency cutoff constrain spending. Free Better Stack checks monitor availability and operator attention; incident and recovery email delivery was exercised. The approved hosting plan has a scheduled retirement after judging, rather than leaving a paid service running indefinitely.

Challenges and what I learned

An exact source quote can coexist with an unsupported sentence. Small live tests exposed stronger assurance wording, incomplete citation coverage and lost approximation qualifiers that the model review did not always flag. I preserved the original result, made seven corrections in a separate cached copy, and kept review warnings visible. All ten sections in the demonstrated edited copy still need review. Proposed objectives and rubrics are identified as proposals.

The video records the actual hosted application and an existing real Nemotron result. It explicitly labels the edited cached copy and does not present cached reuse as a new inference call. An idle interval is cut for length; no generation progress is fabricated.

Validation and accomplishments

Both provider paths have mocked regression coverage. Linux, Apple Silicon Mac and Intel Mac CI exercised tests, builds, packaging and fresh Python-only installation. The actual hosted service was checked for source evidence, editing, exports, session isolation/deletion, restart persistence, zero-cost cached reuse and mobile layouts. Routine tests use mocks; live-provider evidence is reported separately.

Three approved real local/hosted jobs reported 20,932 input and 5,801 output tokens, estimated at $0.0115005 using the published rates. The hosted generation plus review reported $0.0043341 and completed in 15.84 seconds. These are small integration observations, not a quality, cost or latency guarantee.

Existing project and contest work

Easy Street v1.0.0 was released before this adaptation. The October 16 contest changes add the Nebius provider, private judging configuration, sample workflow, abuse and spending controls, deployment and monitoring. The repository records the baseline and changes; earlier features are not presented as newly invented for the adaptation.

What's next

Evaluate more legally usable subjects and improve support warnings without silently adding outside facts or clearing uncertainty. Classroom grading and LMS integration remain future work.

Source, setup and detailed validation: https://github.com/agammann/easy-street/tree/hackathon/nebius-nemotron-2026 Working judging demo: https://easy-street-agammann-2026.fly.dev

Built With

• fastapi
• nebius
• nvidia
• python
• react
• sqlite
• typescript