CITY SCALE · UMD M-TRAIL & MTI · LIVE AT CERRA.CLOUD

CERRA: Auditable Geospatial AI & Traffic Simulation for Public Agencies

Public transportation agencies cannot make signal-timing, work-zone, or freight-corridor decisions from opaque black-box LLM chat replies. Built and operated at the University of Maryland, CERRA pairs real corridor telemetry with a protected LLM-SUMO microscopic simulation bridge — returning observed metrics, simulated counterfactuals, explicit assumptions, and verifiable data sources side by side.

Live Research Platform: cerra.cloud LLM-SUMO Launched Oct 2026 PAR Recommended Practice IBM Heron QPU Benchmark
CERRA geospatial AI platform and LLM-SUMO microscopic simulation architecture
CERRA + LLM-SUMO SYSTEM ARCHITECTURE CLICK IMAGE TO ENLARGE
4-Part Auditable Copilot Contract Observed Data + SUMO Counterfactual + Explicit Assumptions + Sources
LLM-SUMO Protected Execution Bridge Authored by Jihoon Kim; Linux-ported, bridged & operated by Bo Shang
PAR Standard Perceived Acceleration Rate Recommended Practice for heavy-vehicle & corridor stability evaluation
IBM Heron Feasibility-Preserving VRP 100% constraint-valid QAOA routing encoding verified on real QPU hardware

NO HALLUCINATED TRAFFIC ENGINEERING

The Four-Part Auditable Decision Output

CERRA is an AI-driven geospatial research platform built for public agencies at UMD M-TRAIL (under Prof. Xianfeng Yang). Rather than asking a language model to guess queue lengths, CERRA translates natural-language corridor questions into deterministic data queries and sandboxed SUMO simulation runs.

AGENCY ENGINEER QUERY “What happens to I-95 / arterial queue spillback if Lane 3 closes 14:00–17:00?” cerra.cloud/llm-sumo PROTECTED LINUX BRIDGE CERRA + LLM-SUMO • Spatiotemporal SQL/GIS fetch • Parameterized SUMO execution • Schema & bounds guardrail Zero unverified shell access 1. OBSERVED BASELINE: Historical detector speeds, volumes & incidents 2. SUMO COUNTERFACTUAL: Simulated delay, queue length & PAR metrics 3. EXPLICIT ASSUMPTIONS: Demand scaling, car-following & signal phase params 4. VERIFIABLE PROVENANCE: Sensor IDs, timestamps & reproducible config
PLATFORM CORE

CERRA Geospatial Layer

Ingests multi-agency traffic, weather, incident, and infrastructure asset layers into an interactive web map built and operated by Bo Shang at UMD.

DIGITAL TWIN MODULE

LLM-SUMO Integration

Traffic-simulation digital twin authored by M-TRAIL colleague Jihoon Kim; ported to Linux, wired through a protected execution bridge, packaged, and operated in production inside CERRA by Bo Shang (Oct 2026).

STABILITY STANDARD

PAR Recommended Practice

Integrates Perceived Acceleration Rate (PAR) corridor stability diagnostics alongside travel-time and queue metrics for mixed heavy-truck traffic.

NEXT-GEN COMPUTE

Quantum-Safe & QPU Eval

Evaluates post-quantum V2I signal broadcast latency overhead and feasibility-preserving Vehicle Routing Problem (VRP) encodings on IBM Heron hardware.

LIVE DEPLOYMENTS & RIGOROUS BENCHMARKING

Deployed Modules & Hardware Findings

Live Production Research Platform · UMD M-TRAIL

CERRA & LLM-SUMO Web Deployment

Operated at the University of Maryland Department of Civil & Environmental Engineering and Maryland Transportation Institute under Prof. Xianfeng Yang.

Quantum Hardware Evaluation · October 2026

Feasibility-Preserving VRP Encoding on IBM Heron Superconducting QPU

Demonstrated structural route-feasibility guarantees on physical IBM Heron quantum hardware while rigorously benchmarking against classical Genetic Algorithm (GA) and Simulated Annealing (SA) baselines (confirming no quantum runtime speedup over classical heuristics at current NISQ problem scales).