ORChestra: AI Conductor for Operations Research

Describe any operations problem in plain English; ORChestra's AI conductor routes it to one of eight real OR-Tools optimizers over MCP, solves it, proves it optimal, and re-optimizes live.

Comment

About the project (Project Story — Markdown)

🎼 What it is

ORChestra is a chat-native conductor for operations research. You describe a real-world problem in plain English; an AI orchestrator routes it to the right one of eight specialized OR optimizers — each a genuine OR-Tools model exposed as an MCP tool — solves it, and explains the answer. Six of the eight are exact models that report a 0% optimality gap — provably optimal, not a heuristic guess.

💡 Inspiration

Orion is judged by INFORMS — operations-research people — under the banner "Where Operations Research Meets Innovation." We bet the field would be flooded with RAG chatbots, so we went the other way: fuse agentic AI with real optimization. OR has always had a last-mile problem — the models are powerful but locked behind experts and clean data. LLMs are the missing pieces: the layer that turns messy human language into optimizer input, and the interface that lets anyone interrogate a rigorous model. We didn't put a chatbot on top of OR. We gave OR a voice.

🛠️ How we built it

  • 8 OR-Tools optimizers — LifeChain (kidney paired-exchange, CP-SAT cycle/chain packing), FoodRescue (VRP + time windows), ReliefRoute (allocation + fairness floor), ShiftHeal (nurse rostering), ColdChain (cold-chain delivery), FairAid (concave-utility cash allocation), CarbonRoute (multi-objective sourcing), OR-Copilot (knapsack/assignment).
  • MCP layer (FastMCP) — every optimizer is an optimize_* tool; runnable as one aggregated server or eight standalone servers.
  • Orchestrator, two ways — a portable Conductor (Claude routing + keyword fallback + session memory for "what-ifs") and a Hybrid path where Accenture AI Refinery's Distiller drives an MCPClientAgent connected to our MCP server over SSE.
  • Interfaces — a Slack-style web chat (FastAPI) plus a swappable Slack Bolt adapter.
  • Runs with zero credentials; Claude, AI Refinery, and Slack are graceful enhancements.

🎯 The demo beat (LifeChain "Living Optimization")

Optimize a 18-pair kidney pool → 13 transplants, +5 vs a naive matcher, optimality gap 0%. Then say "prioritize the children" → it re-solves live in ~1s: still 13 transplants, but pediatric recipients rise 3 → 5 — a Pareto-improving fairness change, steered in plain English.

🧗 Challenges we ran into

  • AB blood type = universal recipient, which turned a "find an incompatible donor" loop into an infinite loop — fixed with a bounded, realistic crossmatch model.
  • Reproducibility: Python's salted hash() and multi-worker CP-SAT tie-breaking made the "provably optimal" demo flap between runs. We made the crossmatch a stable integer mix and pinned CP-SAT to a single worker + fixed seed — now byte-identical every run.
  • A mislabeled mcp PyPI package; pivoted to the reliable standalone FastMCP.

📚 What we learned

Constraining the LLM to fill typed parameters of pre-validated OR templates (never emit raw solver code) is what makes "talk to the optimizer" safe in front of experts — the model can't hand a judge a wrong model. And honesty scores: we label the 6 exact solvers gap 0 and the 2 routing solvers as metaheuristic, rather than overclaiming.

🏆 Accomplishments

8 real OR models · MCP tool layer · portable and AI Refinery orchestration · web + Slack · 29/29 tests pass · all 8 solve in ~4s from one command.

🚀 What's next

Real EHR/FHIR ingestion for LifeChain, live map/graph visualizations, and a national-pool scale demo.

Built With

  • agentic-ai
  • ai-refinery
  • anthropic-claude
  • constraint-programming
  • cp-sat
  • fastapi
  • fastmcp
  • fintech
  • google-or-tools
  • healthcare
  • llm
  • mcp
  • milp
  • networkx
  • operations-research
  • optimization
  • pydantic
  • python
  • slack
  • slack-bolt
  • sustainability
  • uvicorn
  • vehicle-routing
Share this project:

Updates