FineTuned
Who it is for: A hiring manager or HR coordinator at a small software company or startup who is not an engineer, and who hires junior or intern developers from CVs that list GitHub projects. What goes wrong today: The CV says "Python, SQL, FastAPI" and links a repository, but the hiring manager cannot read the code, so the claim stays unchecked. Asking an engineer to review every repository takes time, so it rarely happens. The interview then relies on generic questions ("Tell me about your project") that a candidate can answer fluently without having built the project, or with AI-prepared answers. Even a coding test does not help much, because it is usually a generic puzzle unrelated to what the candidate claims to have done. What changes with our tool: The manager uploads the CV, picks two skills relevant to the vacancy and gives the candidate's public repository. The tool reads the repository and prepares two things from it. First, interview questions tied to real files and code in that repository. Each question points to a code snippet that the system verifies exists, and questions that cannot be verified are rejected. Second, a short live coding task based on a function from the same repository, for example fixing a bug or handling an edge case. The candidate answers and writes the code in the app during the interview, in person and without AI help. The report then separates what the CV claims, what the repository evidence shows, and what the candidate demonstrated in the interview. A human makes the final decision.
Preliminary score
- Value for the user18.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI22.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing13.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.0 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- Claude Opus 5.5Preliminary70/100
- Strengths
- Your target user is sharply defined: a non-engineer hiring manager at a small company who cannot read the linked repository. The prototype goes well beyond a wrapper. It uses AST-based snippet extraction pinned to a frozen commit, revalidates every code reference the model cites, and drops findings it cannot verify. Your three-column report (Claimed / Repo Evidence / Demonstrated) with no automatic score is a thoughtful, human-in-the-loop design. Your testing write-up is unusually honest: you clearly state that live Groq generation was not re-verified for this version.
- To improve
- Your tests mostly cover plumbing (retries, caching, rate limits) rather than whether the AI output is any good. Run a small evaluation, for example 10 CV/repo pairs with known inflated claims, and report how often ProRed flagged them correctly and where it failed. Back the 80% screening-time cut and the 15 hours/week figure with a timed comparison against a manual review, even with one recruiter. Align your deck with what you actually built: it says Gemini 2.5 Flash, while the README and models say Groq gpt-oss.
- GPT-6 AstraPreliminary75/100
- Strengths
- Your workflow addresses a specific pain for nontechnical hiring managers and separates CV claims, repository evidence, and interview performance without issuing hire/reject scores. Your commit-pinned snippets, reference revalidation, and explicit mock/real separation make the evidence more auditable. Your testing account is specific and candid about simulated checks, provider failures, and the current version's unverified live assessment.
- To improve
- Re-establish your complete assessment on the current Groq build and include a saved example showing inputs, verified references, candidate answers, and model reviews. Benchmark evidence interpretation and answer grading against engineer judgments, including copied-code and missing-public-evidence cases; contract tests alone do not establish assessment accuracy. Validate your claimed screening-time savings against manual review. Update your deck to match the Groq-only implementation and absence of self-practice, and substantiate the $0.02 estimate with current-model token usage and retries.
The team
Fatmakhanim Baghiyeva
data
Zeynalli NasibaLead
developer
Kheyransa Gorchuyeva
data
Alinur Zamanli
developer
AI/ML engineer, Afag Ramazanlı
developer