Third City Analytics
Architecting
the end state.
Systems that show their work
We build analytical systems for operators who need conclusions they can defend — vision pipelines with traceable evidence behind every finding, agent architectures constrained at runtime, and network models for structure that resists clean classification.
Capabilities
Spatial Intelligence
Physical environments, turned into structured, queryable evidence. Our multi-stage vision pipelines run condition assessment, compliance scoring, and capital prioritization across large property portfolios. Every finding links back to traceable visual evidence, so conclusions survive scrutiny from underwriters, auditors, and counterparties. The same systems extend to catastrophe-affected environments, where the data picture is degraded and the cost of delay compounds: compressing the gap between an event and a legible picture of it.
Agentic Systems & Inference‑Time Control
Orchestrated AI for multi-stage analytical work, built around the discipline most deployments skip: runtime control. Enforced structured outputs, deterministic validation that catches hallucinated citations and schema drift before they ship, repair constrained to a fixed catalog of operations, and provenance linking every claim to its source. Agent hierarchies that decompose hard problems and return conclusions — not just data.
Graph Networks & Adversarial Structure
Methods for layered, heterogeneous graphs where the most interesting entities resist clean classification — nodes with probabilistic membership in many communities at once. Mean-field and graphon convergence analysis, multi-member stochastic block models, and behavior-based traversal. Applied to financial network intelligence, supply-chain risk, and entity resolution.
Evaluation & Adversarial Stress Testing
Test-and-evaluation harnesses for AI systems that must be trusted before they are believed. Repeatable evaluation pipelines with deterministic checks, post-hoc validation, and structured failure-mode logging — built to find where agentic systems break under adversarial pressure, before deployment does.
Quantitative Modeling & Causal Inference
Econometric and formal modeling where clean experiments are unavailable. Quasi-experimental design, agent-based simulation, and formal specification under incomplete information — the measurement backbone that says whether a system actually worked.
Work
with us.
If you have a problem where the standard toolbox does not reach, we would like to hear about it.