HackInsight: Data-to-Decision

Build an automated pipeline, and an agentic AI system on top of it, in 72 hours.

Mode: Online | Team size: 4 students

The challenge

Choose a real problem that matters to people in Pakistan. Build a dataset strong enough for a peer-reviewed journal: traceable, licensed, documented, checked and reproducible. Automate the pipeline, then add agentic and generative AI that monitors quality and explains results without inventing anything.

About datasets

No dataset is provided, and teams do not create data themselves. Each team gathers its data from credible, original publishers during the event and declares it by Mon 19 Oct, 12:00 PM PKT.

- Use at least 2 independent sources, including at least 1 credible source not on the track list.

- Name the original publisher, URL or DOI and licence for every source.

- No synthetic or AI-generated data for training or evaluation.

- No personal or identifiable data.

- Respect each site's terms of use and robots.txt.

What to build

1. Data acquisition and warehouse: raw, cleaned and processed layers rebuilt with one command (make data).

2. Data Quality Agent: runs validation tools, decides to fix, fill or flag, and explains each decision.

3. Model: beats a declared baseline, with uncertainty and errors by group.

4. Insight Agent (RAG): cited explanations for the primary user; says "I don't know" when unsupported.

5. Automated pipeline and dashboard: hourly flow plus a Hugging Face Space showing data freshness, quality status and last successful run.

Tracks

- Climate and heat-health

- Air quality

- Energy demand

- Public health

- Food security

- Education

- Information integrity

- Urban mobility

Teams may propose their own problem if the data meets the same standard.

Judging (100 points)

- Automatic checks (baseline improvement, reproducibility, honesty): 40

- Dataset selection, integrity and strength: 20

- Agentic design and automation: 15

- Scientific rigour: 10

- Usefulness and responsible AI: 10

- Viva and demo: 5

Pakistan Impact Bonus: up to 10 extra points. Ties go to the higher bonus. The bonus cannot cancel a failed integrity check.

Bonus (judged)

Data Enrichment Agent, Urdu output, Supervisor agent, drift monitor, human-in-the-loop approval, 2-page data paper draft, Change Request 2.

Research award

The strongest dataset and evaluation may be invited to develop into a journal submission with faculty mentors.

• Automatic checks
  Baseline improvement, reproducibility and honesty of results
• Dataset selection, integrity and strength
  Traceable, licensed, documented, checked and reproducible data from 2 or more independent sources
• Agentic design and automation
  Data Quality Agent, Insight Agent (RAG) and the hourly pipeline
• Scientific rigour
  Sound baseline, uncertainty and errors by group