How it's built
Deterministic where it counts. Generative where it helps.
"AI powered" tells you nothing. Here is what we actually built, and why a bid director, an operator or an authority can trust it.
Quick answer
YourTransportBrain puts AI agents on a transport ontology: tenders, claims, outcomes, contracts and KPIs as typed, linked objects. Rules and maths make the decisions that have to be right; Claude models draft and explain. Every claim is released, hedged or withheld on its evidence, every agent handoff is a recorded event, and no agent ever checks its own work.
The principles
Ten things we built on purpose.
- 01
Agents on an ontology, not a pile of PDFs
Tenders, claims, outcomes, contracts and KPIs are typed, linked objects with a state machine. The agents reason over the objects, not over loose documents.
- 02
Deterministic where it counts. Generative where it helps.
Rules and maths decide. The language model drafts and explains. Classification, outcomes and checks never rest on a model's say-so alone.
- 03
The model is one voter. Never the judge.
Every document is classified from its metadata, deterministic rules and a model vote. Confidence is anchored to evidence, never to the model's opinion of itself.
- 04
Release, hedge or abstain
Every claim carries a truth state. Strong evidence is released, thin evidence is hedged, no evidence means it says so. A weak claim degrades; the answer does not.
- 05
Two corpora, one retrieval
A shared public transport corpus and your private corpus are searched together and never mixed. Separate schema per customer, row-level security, no cross-tenant flow.
- 06
Event-sourced agents
Every agent decision writes a typed event that every other agent can read. Handoffs travel on that spine, so every handoff has a receipt.
- 07
Outcomes triangulated, forecasts scored
A win or loss commits only on two or more corroborating sources. Every prediction carries a resolve-by date and is marked right or wrong when it lands.
- 08
The doer never checks itself
Every agent has a separate checker: deterministic checks first, a different model second. Golden tests run against the live AI every night.
- 09
The right model for each job
Claude Opus 4.8 drafts and reasons. Claude Haiku 4.5 classifies and extracts at volume. OpenAI embeddings index. Postgres with pgvector holds it.
- 10
Intent in, decisions out
Orchestration lives in the platform, not the screen. You see decisions and exceptions with the evidence attached, then approve, tweak or send back.
The stack
Seven layers, one brain.
- 07Agent outputDecisions and drafts, each with its evidence trail, waiting for a person.
- 06Company layerYour private data, walled: bids, evidence, contracts, readings.
- 05DatabasePostgres, a separate schema per customer, row-level security.
- 04OntologyTenders, claims, outcomes, contracts and KPIs as typed, linked objects.
- 03ConnectEvery object joined to the buyers, operators, routes and notices it touches.
- 02IndexEmbeddings and search over the public corpus and yours, kept apart.
- 01IngestLive public feeds and your documents, every one fingerprinted and checked before it counts.
We run on it
Our own company runs on agents.
25 agents run economAIcs today. They brief the CEO each morning, watch every system's heartbeat, test the AI against golden cases every night and harvest the market. The same patterns go into every agent we build for you: a separate checker, deterministic checks first, one alert per fault, outside content quarantined before an agent reads it, and a person approving every release.
FAQ
Which AI models does economAIcs use?
Claude Opus 4.8 drafts and reasons. Claude Haiku 4.5 classifies and extracts at volume. OpenAI embeddings index the corpus, and Postgres with pgvector holds it. Each model does the job it is best at.
Does the AI make things up?
Every claim carries a truth state: released when the evidence is strong, hedged when it is thin, withheld when there is none. It says it does not know before it guesses, and every answer cites its source.
Is my data used to train models?
No. Our AI providers are contractually barred from training on it. Your data sits in its own schema, behind row-level security, and never mixes with anyone else's.
What does deterministic mean here?
Where a decision has to be right, rules and evidence have the final say, not a language model: classifying documents, confirming outcomes, scoring forecasts, checking every draft before it reaches you. The model drafts and explains around those facts.
Who checks the agents?
Never the agent itself. Every agent has a separate checker, deterministic checks first and a different model second, and golden tests run against the live AI every night.