FRIDA — Autonomous Strategic Urban Intelligence

FRIDA notices changes nobody thought to ask about, connects signals, challenges assumptions, and surfaces what decision-makers should see before it becomes obvious. Nobody asked her. She noticed.

Comment

Inspiration

Cities rarely announce that something important is beginning to change.

The first signs appear separately. A development advances. Population projections shift. Transport conditions change. Infrastructure context adds another clue. Public information suggests that something deserves a closer look.

Each fact may be ordinary on its own.

Together, they may be telling us something.

Most urban intelligence systems wait for someone to ask a question. But strategic change often begins before anyone knows which question to ask.

That is where FRIDA begins.

FRIDA notices changes nobody thought to ask about, connects signals, challenges assumptions, and surfaces what decision-makers should see before it becomes obvious.

Nobody asked her. She noticed.

What it does

FRIDA is an autonomous strategic urban intelligence system.

She continuously observes authorized evidence about the city, preserves its provenance and temporal state, and determines whether new information deserves strategic attention.

She does not treat every new record as an alert.

Instead, FRIDA performs a governed multi-step workflow:

Observe → Appraise → Investigate when warranted → Look ahead → Govern → Brief

A bounded cognitive First Appraisal asks whether an observation may have strategic significance and identifies what evidence is still missing.

If further investigation is justified, FRIDA conducts bounded research over authorized evidence. She can then perform Foresight analysis: not prediction, but a disciplined exploration of plausible trajectories, leading indicators, uncertainties, and intervention or opportunity windows.

The result is a durable Strategic Briefing, not a chat response.

The current London demonstration shows FRIDA evaluating development activity, transport context, population projections, environmental and resilience information, and aggregate place-based safety context.

Her current persisted London finding is an advisory Yellow — POSSIBLE / INTERVENTION_OPPORTUNITY. FRIDA explicitly preserves what is still unknown instead of converting incomplete evidence into false certainty.

The public demonstration is deliberately read-only. Judges can move from the executive briefing to the evidence, execution transparency, and Raw Governed Record behind it.

How we built it

FRIDA was designed as an autonomous multi-step agentic workflow under deterministic governance, rather than as a chatbot with a longer prompt.

Her central design principle is:

Determinism establishes what is true and what is allowed. Cognition explores what it might mean.

The deterministic layer controls:

  • source authorization;
  • provenance;
  • normalization;
  • fingerprints and deduplication;
  • temporal integrity;
  • privacy;
  • client isolation;
  • schema validation;
  • evidence scope;
  • spend limits;
  • governed state transitions;
  • executive posture;
  • persistence.

Gemini is used only where cognition adds value:

  • First Appraisal
  • Bounded Research synthesis
  • Foresight
  • Executive Brief wording

The model may write the sentence.

It may not choose the posture.

The submitted production runtime uses:

  • Gemini 3.6 Flash
  • Google GenAI SDK (google-genai)
  • Vertex AI
  • Google Cloud Run
  • Cloud SQL for PostgreSQL
  • Python with structured-output contracts and deterministic validation

FRIDA is deployed publicly on Cloud Run and persists governed artifacts in Cloud SQL.

The London evidence fabric currently includes authorized public information from:

  • Planning London Datahub
  • Transport for London
  • Environment Agency
  • Greater London Authority housing-led population projections
  • Metropolitan Police Service aggregate borough-level context

Official announcements and targeted urban news can act as governed clues, but:

A clue is not a conclusion.

Architecturally, FRIDA follows a pragmatic combination of Hexagonal / Ports-and-Adapters, bounded domain contexts, and an append-only auditable ledger.

Strategic memory and time

Cities evolve over years, while hackathon demonstrations last minutes.

FRIDA therefore separates evidence time from generation time.

Historical Strategic Briefings can be generated using FRIDA's current intelligence while deterministically restricting her field of view to evidence that was actually knowable at a declared historical cutoff.

The submitted build includes governed historical London briefings with evidence cutoffs in 2006 and 2021, alongside the current Strategic Brief.

This is the foundation for what we call Historical Onboarding: allowing FRIDA to learn a city's governed temporal history before entering live observation.

Our rule is simple:

We accelerate time — not evidence, truth or intelligence.

Challenges we ran into

The hardest part of FRIDA was not making Gemini produce plausible interpretations.

It was deciding when the system was allowed to believe them.

Urban data is heterogeneous. Sources have different update cycles, geographic precision, and original purposes.

Population projections are not observed demand.

A transport delay is not proof of structural capacity pressure.

Aggregate safety statistics cannot establish local causality.

An environmental source should not be forced into an unrelated hypothesis merely because it is available.

We repeatedly encountered plausible relationships that the evidence did not yet justify.

That led us to make uncertainty and restraint first-class product behaviors.

FRIDA can decide that something deserves further investigation.

She can also decide that it does not.

And a legitimate Green result is not “everything is fine.” It means:

No material strategic change was detected in the monitored evidence during that period.

Another challenge was temporal integrity. Historical analysis is easy if a model is accidentally allowed to know the future.

We therefore enforce historical evidence cutoffs deterministically before model invocation, rather than relying on prompts asking the model to ignore later knowledge.

Accomplishments that we're proud of

We are especially proud that FRIDA became a system that completes strategic work autonomously instead of waiting for a sequence of human prompts.

She can:

  • monitor authorized urban evidence;
  • decide whether something deserves cognitive attention;
  • formulate a better strategic question;
  • identify missing evidence;
  • conduct bounded follow-up research;
  • reject unsupported relationships;
  • explore plausible future implications;
  • preserve uncertainty;
  • deterministically govern what can be claimed;
  • publish a persistent Strategic Briefing.

The current London briefing was not manually written and then dressed as an AI result.

It emerged from the governed workflow and remains connected to the evidence that produced it.

FRIDA also exposes multiple levels of transparency:

  • executive presentation;
  • Strategic Briefing history;
  • Under the Hood execution views;
  • Raw Governed Records.

And perhaps the accomplishment we value most is that FRIDA does not need someone to know the right question first.

What we learned

We learned that useful autonomy is not the absence of constraints.

It is curiosity with disciplined boundaries.

We learned that deterministic systems and generative AI do not need to compete for authority. Each is better at a different job.

Determinism is excellent at establishing truth boundaries, permissions, state, and reproducibility.

Cognition is excellent at asking what evidence might mean and what question should come next.

Putting those two capabilities together produced a system that became more useful as we allowed it to be curious — without allowing curiosity to become certainty.

One First Appraisal captured this particularly well: when FRIDA was allowed to consider whether something might matter, her first response was not a dramatic conclusion.

She asked for more evidence.

What's next for FRIDA

The public London demonstration intentionally uses a limited set of authorized public sources.

A city government could provide richer, higher-resolution, and more timely authorized evidence: infrastructure capacity, detailed mobility loads, development completion and occupancy, service demand, resilience information, and additional institutional datasets.

That would improve the precision and completeness of FRIDA's assessments without changing the governing principles behind them.

Next, we want to expand:

  • source discovery and authorization;
  • richer temporal pattern memory;
  • broader Historical Onboarding;
  • additional specialized cognitive capabilities when strategically warranted;
  • notification and briefing delivery workflows;
  • deployment across additional cities and urban systems.

The larger opportunity is not simply adding more data.

It is creating strategic intelligence that notices when the future has already started — while still being able to show exactly why it thinks so.

Nobody asked her. She noticed.

Built With

Share this project:

Updates

Submission history