convez.ai
SkillBusca & PesquisaIntermediário
LeSingh1

Reverse Engineer de Hackathons

LeSingh1/hackathon-rce

Descubra padrões de projetos vencedores e estratégias para sua hackathon

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que indexa 1.319 projetos vencedores de hackathons em busca vetorial (ChromaDB) e extrai padrões táticos de sucesso. Dado os detalhes de um hackathon (tema, patrocinadores, trilhas), retorna os projetos mais relevantes do passado e ideias personalizadas com alta probabilidade de vencer.

Instalação

$npx convez-ai LeSingh1/hackathon-rce

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Identificar ideias de projetos que já venceram hackathons similares
  • Descobrir padrões comuns em projetos premiados por trilha/tema
  • Gerar estratégias de participação baseadas em dados históricos reais

Veredito

Ferramenta única que transforma dados brutos de 1.319 vencedores em insights estratégicos acionáveis. Extremamente útil para participantes que querem otimizar chances reais baseado em histórico, não em genéricos — o único risco é depender demais de padrões passados em temas emergentes.

README

A tradução para o português deste README ainda está na fila.

Hackathon Reverse Engineer (hackathon-rce)

Reverse-engineer what wins hackathons. This tool indexes 1,319 winning hackathon projects scraped from Devpost into a local vector database. Give it details about a hackathon you're about to enter — name, theme, sponsors, tracks — and it surfaces the most relevant past winners, then derives the concrete patterns that won and tailored project ideas built to win your target hackathon.

Built to be invoked as a Claude Code skill: /hackathon-rce.


Why this exists

Most "hackathon advice" is generic ("solve a real problem", "make a good demo"). This tool does the opposite: it pattern-matches against what actually won across 408 distinct hackathons and 884 distinct judging tracks, then tells you what to build.

StatValue
Winning projects indexed1,319
Distinct hackathons covered408
Distinct judging tracks884
Projects using AI agents659
Years spanned2014–2026 (bulk 2023–2026)
With verified Devpost source URLs1,280

Domains covered include Healthcare, Education, FinTech, Climate/Sustainability, Web3, Accessibility, Hardware/IoT, AR/VR, Voice AI, AI Agents, Social Impact, Cybersecurity, and more.


How it works

                       ┌──────────────────────┐
   Hackathon details   │   /hackathon-rce      │
   (name, theme,  ───► │   Claude Code skill   │
    sponsors, tracks)  └──────────┬───────────┘
                                  │ runs
                                  ▼
                       ┌──────────────────────┐
                       │      query.py        │  semantic search
                       │  (retrieval only)    │
                       └──────────┬───────────┘
                                  │ queries
                                  ▼
                       ┌──────────────────────┐
                       │      ChromaDB        │  all-MiniLM-L6-v2
                       │  hackathon_winners   │  cosine similarity
                       │   1,319 documents    │
                       └──────────┬───────────┘
                                  │ returns top-12 winners as JSON
                                  ▼
                       ┌──────────────────────┐
                       │   Claude (native)    │  analyzes JSON →
                       │  PATTERNS + IDEAS    │  tactical strategy
                       └──────────────────────┘

The retrieval step (query.py) does only vector search and emits JSON — no LLM call, so it needs no API key. Claude does the analysis natively inside the skill. See ARCHITECTURE.md for why.


Setup

bash setup.sh

This creates an isolated .venv, installs dependencies (pinning a working httpx/chromadb), and builds the vector index from data/winning_projects.json into a persistent ChromaDB at data/chroma_db/.

Why a venv? The query engine imports chromadb, which is sensitive to a broken httpx on sys.path (e.g. a stray editable install). The venv pins known-good versions so the engine can't be broken by unrelated global-env changes. Always invoke via .venv/bin/python.

Manual setup, if you prefer:

pip install -r requirements.txt
python src/db.py

Usage

As a Claude Code skill (recommended)

The skill lives at ~/.claude/skills/hackathon-rce/SKILL.md. Just type:

/hackathon-rce

Provide your hackathon's name, theme, sponsors, and tracks. Claude runs the query and returns two sections: PATTERNS FROM PAST WINNERS and PROJECT IDEAS.

As a CLI (raw retrieval)

python src/query.py "MLH ShellHacks 2026" "AI for social good" "OpenAI,MongoDB,ElevenLabs" "Healthcare,Social Good"

Prints JSON with the top-12 most similar winners. Pass "" for any field you don't have.

As a standalone analyzer (optional, needs API key)

export ANTHROPIC_API_KEY=sk-ant-...
python src/analyzer.py "MLH ShellHacks 2026" "AI for social good" "OpenAI,MongoDB" "Healthcare,Social Good"

Calls the Anthropic API directly and prints the full patterns + ideas analysis. Without the key it falls back to printing the retrieved winners for manual analysis.


Building / refreshing the dataset

The dataset was built by scraping Devpost winner galleries and search results via Exa (Devpost sits behind Cloudflare, so direct requests scraping is blocked). The pipeline:

  1. Discover — parallel Exa searches across ~90 themed queries + 50 hackathon winner galleries → collect devpost.com/software/* URLs
  2. Extract — fetch each project page, parse name / hackathon / prize / tracks / tech stack / sponsors / metrics into structured records
  3. Merge — src/merge_batches.py dedupes (by id, URL, and name), validates, and writes data/winning_projects.json
  4. Index — src/db.py rebuilds the ChromaDB index

To re-merge extraction batches and reindex:

python src/merge_batches.py   # merges /tmp/hrc_batch_*.json into the dataset
python src/db.py              # force-reindexes ChromaDB

src/scraper.py contains a direct-HTTP scraper kept for reference; the Exa-based pipeline is what produced the current dataset.


Project layout

hackathon-rce/
├── data/
│   ├── winning_projects.json   # 1,319 curated winning projects (source of truth)
│   └── chroma_db/              # persistent vector index (gitignored, rebuild via db.py)
├── src/
│   ├── db.py                   # ChromaDB indexing + semantic retrieval
│   ├── query.py                # retrieval-only CLI → JSON (used by the skill)
│   ├── analyzer.py             # optional standalone Anthropic-API analyzer
│   ├── main.py                 # interactive CLI wrapper
│   ├── scraper.py              # direct-HTTP Devpost scraper (reference)
│   └── merge_batches.py        # dedup + validate + merge extraction batches
├── tests/                      # pytest suite (26 tests)
├── requirements.txt
├── pytest.ini
├── README.md
└── ARCHITECTURE.md

Testing

python -m pytest

26 tests cover the indexing layer, the merge/validation pipeline, the query JSON contract, and integrity of the production dataset (uniqueness, required fields, track diversity, plausible years).


Data schema

Each project in winning_projects.json:

{
  "id": "salus",                       // slug, unique
  "name": "Salus",
  "hackathon": "UofTHacks 13",
  "year": 2025,
  "prize": "Best Use of MongoDB Atlas",
  "tracks": ["Healthcare AI", "Best Use of MongoDB Atlas"],
  "description": "...",                // what it does + key innovation
  "inspiration": "...",                // the problem that inspired it
  "tech_stack": ["FastAPI", "Next.js", "MongoDB Atlas"],
  "sponsors_used": ["MongoDB"],
  "uses_agents": true,
  "agent_count": 3,
  "key_factors": ["...", "..."],       // why it won
  "metrics": ["97% accuracy", "<2s latency"],
  "approach": "...",                   // core architecture, one line
  "judge_appeal": "...",               // what makes judges pick it
  "source_url": "https://devpost.com/software/salus"
}
Ver o original no GitHub

Bom para

Relacionados