remapd

Improving equity by bringing transparency to congressional district boundaries. Built for stakeholders who are traditionally underrepresented in the creation of legislation that directly affects them.

Comment

Inspiration

We were motivated by the ongoing harms of gerrymandering and the long legacy of redlining-driven inequity in U.S. representation. Redistricting often happens in processes that are hard for the public to inspect, which creates a gap between who is affected and who can actually evaluate the maps.

Our goal with remapd is to make districting more transparent, measurable, and fairness-oriented so that marginalized communities can respond with evidence, not just suspicion.

What it does

remapd is a redistricting analysis platform that:

  • Optimizes county-to-district plans with a weighted multi-objective function.
  • Scores plans on fairness, population balance, compactness proxy, and voting-rights proxy.
  • Compares optimized outputs to a baseline for clear, quantitative improvement tracking.
  • Fetches Census-grounded data through a tool layer to support factual analysis.
  • Produces structured, interpretable, natural-language outputs through a multi-agent LangGraph backend.

How we built it

  • Frontend: Next.js + React + TypeScript + shadcn/ui, with D3/TopoJSON maps and Recharts metrics.
  • Backend: FastAPI + Pydantic + NumPy for optimization APIs and scoring.
  • Optimizer: Reward-guided local search (simulated annealing style) over county assignments.
  • Agents: LangGraph orchestration with Engine, Civil Rights, Legislative, and summary-style outputs.
  • Data grounding: MCP-style tools for Census fetch + provenance/audit signals.
  • LLM layer: Anthropic Claude for non-technical structured explanations.

Challenges we ran into

  • Hallucination risk: Preventing unsupported claims in agent output, which pushed us to ground with MCP-style Census tools.
  • Data/API reliability: Handling key setup, API errors, and consistent runtime behavior.
  • Algorithm scope under time constraints: We originally aimed for PPO-based RL training, but given hackathon timelines and training overhead, we shifted to a more efficient reward-guided objective search approach to deliver a reliable end-to-end system.
  • Audience refinement: We initially centered policymakers, but through team iteration shifted focus toward advocacy groups, constituents, and public-interest users who need actionable evidence.
  • Reward design and tuning: Balancing multiple objectives so no single metric dominated.

Accomplishments that we're proud of

  • Built a full pipeline from optimization to plain-language explanation.
  • Delivered baseline-vs-optimized transparency with metric-level detail.
  • Implemented an easy-to-use interface for non-technical users.
  • Grounded analysis in real Census-backed data.
  • Created a workflow that helps civic actors move from critique to evidence-backed alternatives.

What we learned

  • Explainability and trust are core requirements in civic AI.
  • Grounding external facts is essential in high-stakes policy contexts.
  • A simpler optimization approach can be more practical than heavy training under hackathon constraints.
  • Product framing matters: tools for advocacy and constituents need clarity, accessibility, and defensible outputs.

What's next for remapd

  • Work with various advocacy groups and communities that are traditionally underrepresented to draft concrete, real policies that provide fairer district maps.
  • Strengthen legal/policy context for LangGraph agent outputs using a RAG-support legal database, to further reduce hallucinations.
  • Track real-world impact through simulation-based methods, such as creating a World Model for legislation impact.

Built With

Share this project:

Updates