🚀 Context
Hacknation — a hackathon organized by GovTech Polska. 1500+ participants, 480k PLN prize pool, 24 hours for a working public-administration solution.
- Team: a “retired” developer (4 years away from coding) + non-developers using nocode and LLMs
- Core thesis: AI is the equalizer — a tool that lets a team with less coding experience compete with professional dev teams
- AI was used to pick the task — analyzing challenges against the team’s competencies
🗒️ Problem — Mrs. Zosia and Thousands of Excels
The public-administration budgeting process is built on the manual exchange of hundreds of thousands of Excel files:
- Start (bottom): a clerk manually types budget data into Excel
- Escalation (top): the file travels: Municipality → City → Voivodeship → Ministry of Finance
- Consolidation: a dedicated unit at the ministry merges the data (often by hand)
- Decision and return (bottom): budget limits travel back the same path with arbitrary cuts — nobody can explain why
Solution: Cyfrowy Budżet — a centralized web app:
- One source of truth — all budget items in one system
- Transparency — comments and discussion on items inside the system instead of email
- Approval workflow — a simplified approval and consolidation process
🛠️ Stack
| Layer | Technology |
|---|---|
| Frontend | React, TypeScript |
| Backend | Supabase |
| Presentation | Video generated in HiGen |
💰 Token Usage
- Paweł and Kuba: Antigravity (Gemini Pro / Claude 4.5) — burned the entire weekly token limit in ~15h
- Justyna: Bolt (Claude Code) — a record 18 million tokens
Work ran non-stop for 24h, 2-3h of sleep. Initially everyone built separate pieces of code ad hoc. The turn came after consolidating around Justyna’s most advanced prototype.
⚠️ What Went Wrong
Final score: 2.15 / 5 points — didn’t make the finals.
- The tech worked, the presentation was great
- We lacked validation — the team had no access to a practitioner (a clerk working with the budget daily)
- The mentor assigned to the task wasn’t a domain expert
- The system might have been completely “off” — disconnected from the realities of public administration
🤖 AI Limitations
- Code often didn’t work — solutions looked correct but fell apart at runtime
- Hallucinations — proposed libraries didn’t exist, logic pulled out of thin air
- Needs hand-holding — precise prompting and continuous course correction
- Blockers — a bug in the current user filters the model couldn’t diagnose; we had to fall back to manual code reading and debugging
AI is a powerful force multiplier but not a magic wand. Without technical skills and critical thinking — you’ll get stuck halfway.
☘️ Key Lesson — Validation > Technology
Even the best code won’t save a solution that doesn’t address a real user need.
Teams that came with ready components and better business analysis won. The “wing it” approach is romantic but loses against preparation.
Plan for the next hackathon:
- Task selection — team roles, task scraping, AI analysis (this stage was OK)
- Business analysis (this is where we failed) — AS-IS/TO-BE map, User Stories, PRD, SRS, skills for agents
- Development — ready boilerplate, iterative development, automated tests, continuous Code Review
- Documentation and verification — security review, performance review
📒 Summary
- AI lets you do things that were impossible a year ago — a small team built a working web app in 24h
- Technology is secondary to understanding the problem
- AI is the future of programming, but a human has to be the pilot who knows where they’re flying
- Agentic Coding and tools like Claude Code dramatically lower the barrier to entry, but they don’t eliminate the need for technical fundamentals and business analysis