Inspiration
Feedback to a personal agent often loses its context. A request to keep exact numbers for one report can become a global preference, while a one-task exception can linger. I wanted a useful work agent whose corrections have visible consequences before they become memory.
What it does
Akashic Rehearsal produces an operations brief and launch copy from two synthetic projects. A correction retains its original wording and source. You review its project or one-task scope, optional expiry and conflict handling, then rehearse the next artifact before adopting it. The preview shows the before-and-after work and checks whether the other project keeps its own policy.
Adopted rules, retired versions and consumed task exceptions remain in a correction ledger. Every saved artifact records the policy version, applied rules and generation mode. You can export the workspace for inspection.
How I built it
The responsive English web interface uses JavaScript, HTML and CSS. A Python standard-library service and SQLite store isolated browser workspaces with atomic policy updates. Origin checks, HttpOnly cookies and stale-preview guards protect state changes. A server-side model adapter supports correction extraction and artifact generation, with separate local, Nebius and deterministic offline modes. Failures never silently switch modes.
Challenges and lessons
A real local model sometimes rounded figures, omitted a source identifier or invented trend language. I preserved its raw text, tightened extraction and numeric instructions, and added limited output warnings. The application can append a separately labelled list of provided source IDs. That list is not per-claim attribution, and passing fixture checks is not comprehensive factual verification.
What works today
The public review app supports correction proposals, explicit adoption, actual generated work, one-task exceptions, expiry, version history and cross-project policy isolation. Thirty unit and HTTP API tests pass, and the public repository passes CI on Python 3.11 and 3.12. Local-model generation and the public HTTPS workflow have been exercised. All examples are synthetic; no company records or private personal memory are included.
Current status and next steps
This is the working review build prepared for the Nebius Personal AI track. The NVIDIA-on-Nebius adapter is implemented, but the required sponsor-endpoint execution has not yet been verified. The next step is to connect the eligible NVIDIA model, record that real execution and expand evaluation beyond the two designed sample tasks. I do not claim learned personal weights or measured personalisation gains.
Links
Working app: https://knowledge.openakashic.com/hackathons/akashic/ MIT source: https://github.com/SZARA2/hackathon-2026-akashic-rehearsal
Log in or sign up for Devpost to join the conversation.