Inspiration

A job search often becomes disconnected links, rewritten documents and email threads. A match score alone does not explain whether a candidate's experience supports a requirement. We built Rolemorrow to preserve the facts behind a decision and keep the candidate in control of preparation and communication.

What it does

Rolemorrow connects a master career profile with vacancy discovery, requirement-by-requirement evidence review, saved jobs, application materials and correspondence. Candidates can inspect profile quotations, retain uncertainty, record a next action, and save a CV, a separate cover letter and screening answers in a versioned packet.

The video and downloadable Docker demo use fictional data and deterministic providers. A separate hosted jury account contains only fictional candidate data. Real email requires a confirmed recipient and saved content; replies can create in-app notifications. No employer messages were sent during the jury check. The controlled test email was associated manually because fictional fixture jobs deliberately cannot receive automatic application matching.

How we built it

JavaScript and Node.js power the browser workspace and persistent state. NVIDIA Nemotron models run through Nebius Token Factory: Nano extracts vacancy requirements, while Super supports assessment and review. Tavily supports discovery. Local validation checks quotations, fact references, project scope and metrics around generated output. The frontend is not hosted on Nebius AI Cloud.

Account, private-storage, audit, import, MCP and mail adapters are included in source. SQLite supports the default local catalog; PostgreSQL has a separate implemented and tested path. We do not claim that the hosted service has migrated to PostgreSQL. The default judge build needs no login, API keys or payment and keeps live providers and mail disabled.

Challenges

A number can be correct but belong to the wrong project. A plausible sentence can introduce an unsupported causal claim. We separate generation from validation and preserve refusals rather than presenting valid JSON as a finished CV.

AI-assisted document adaptation remains a quality-limited beta. In the single authorized hosted jury evaluation on October 29, an answer omitted metrics available in the fictional profile. Validation refused the draft and the prior manual packet remained unchanged. We did not retry or weaken the check. The saved manual packet is not presented as AI-generated.

Accomplishments

The reviewed public release passed 679 deterministic/integration tests with zero skips or failures, all 15 browser programs, a separate two-account isolation walkthrough, and clean Docker build checks. An anonymously downloaded archive matched all 191 reviewed files. These are bounded reproducibility checks, not proof of universal model correctness.

The hosted jury check verified registration, email verification, fresh password login, fictional profile, evidence review, packet persistence, a mobile viewport, and receipt of a controlled email with a persistent read/unread notification. No owner's personal workspace was accessed. Actual hosted generation limitations are disclosed above; the hosted build and downloadable source revision are distinct.

What we learned

Source references, uncertainty, project boundaries and recovery need explicit controls. Deterministic tests and real-model evaluations establish different things. Passing tests cannot certify the factual quality of every generated document.

What's next

Improve source-grounded adaptation and diagnostics, expand verified vacancy coverage, and continue user evaluation. Daily personal digests and product analytics remain planned, not demonstrated features.

Development-period disclosure

The October 18 prototype was extended with application workflows, imports, auditability and stricter document validation. The October 29 public-safe release replaces private identity and career-derived fixtures with wholly fictional examples and excludes runtime data and credentials. The entrant must confirm eligibility, contribution and any pre-existing components before final submission.

Built With

Share this project:

Updates

Submission history