Teamup
Founders and hackathon participants in Azerbaijan who need teammates with the skills they lack, for example a developer with a B2B idea who needs a sales co-founder. Today they post in Telegram groups or get put into random teams: a team ends up with four developers and no designer, keyword search only matches people who look alike, and gaps in hours or goals show up months later. Hackathon organizers have no simple way to give every team the roles it needs.
Preliminary score
- Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI19.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing15.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility9.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- GPT-6 AstraPreliminary70/100
- Strengths
- Your complementarity-first workflow addresses a concrete problem for founders and hackathon organizers, with measurable role-balance outcomes. Your structured scoring, profile review, and reason-grounded explanations make recommendations inspectable rather than opaque. Your reported tests include useful baselines and specific failures, and you clearly disclose the offline extractor and unverified labels.
- To improve
- Your next evaluation should use live LLM outputs and independently reviewed labels; the current results do not establish the AI pipeline's performance. Your ranking needs recalibration around availability and goals, given the disagreements in 10 of 15 scenarios. Your organizer pilot should compare against manual team formation, not just random assignment, and track accepted introductions and sustained participation. Your feasibility case would benefit from per-profile and per-event running costs plus a concrete plan to recruit consenting users beyond synthetic profiles.
- Claude Opus 5.5Preliminary71/100
- Strengths
- Your problem framing is specific: a hackathon team with four developers and no designer, or a B2B founder who needs a sales co-founder. You chose a measurable outcome, the share of teams missing a required role. Your architecture separates responsibilities well: the LLM handles multilingual extraction and grounded explanations, while deterministic scoring with reason codes keeps rankings transparent and testable. Your Quality Lab is unusually thorough and honest, covering a BM25 and random baseline, counterfactual fairness pairs, robustness cases including prompt injection in AZ/RU/EN, and concrete failures such as "full-time, but only weekends" being read as full_time.
- To improve
- Your published numbers come from the offline heuristic, not Claude, and the gold labels are synthetic with 0/48 human-reviewed, so the actual AI contribution is still unmeasured. Run `pnpm eval` with a live key and show the LLM against the heuristic side by side. Even after your optimisation, 67% of hackathon teams still miss a required role, and the ranking disagreed with human labels in 10 of 15 scenarios. Analyse those cases and tune the time and goal weights. Add concrete per-profile API cost figures and a named pilot, such as a real Baku event with an organizer, to make the adoption path tangible.
The team
Ravan Guliyev
designer
Tariyel Jabrayilli
developer
Turgut Abbasli
data
Nihad Eminov
developer
Adam HasanovLead
product