Whose Judgement Will Cities Learn From?

 Over the past few weeks, I’ve been working through a simulation of Kochi’s development dynamics - trying to understand how policies like TDR, FSI, and conservation incentives actually interact in practice. Check them out if you are interested.

Run the simulation (browser, no install)
https://ajithvyas.github.io/kochi-tdr-model/TDR_Kochi_Baseline_v05a_web.html

→ Click Setup → GO and let it run (~200 ticks)

Interactive policy dashboard
https://kochi-tdr-sandbox.streamlit.app
→ Test different policy combinations and see how outcomes shift

Full paper (method + findings)

https://zenodo.org/records/19249811

All that work led me to a broader question: not just how we simulate cities, but how these simulations begin to feed into decision-making, and increasingly, into AI systems that learn from such data. This short note is an attempt to think through that shift, and what it might mean for urban designers.

Whose Judgement Will Cities Learn From?

From simulation to decision support

Ajith Vyas Venugopalan, Architect, Urban Designer, Urban Researcher

30 Mar 2026

One of the more interesting properties of an agent-based model is that it doesn't produce a single answer. It produces a landscape of possible outcomes, and that landscape itself is information.

Rather than running every combination manually, you define key variables with meaningful threshold values. Together, these form a behaviour space: a structured map of how the system responds under different policy conditions. In this work, that space is visualised through the dashboard. But it can also be exported as synthetic datasets - each scenario becoming a structured input-output pair, with policy configurations as inputs and spatial or economic outcomes as outputs.

This opens two directions worth taking seriously.

The first is econometric. These datasets allow us to quantify how strongly each variable drives outcomes - moving the model from exploration toward genuine decision support. In the Indian context, this matters in a specific way: our planning systems are full of policies that exist on paper but have never been empirically tested against each other. FSI bonuses, TDR, heritage overlays, CRZ buffers - they interact constantly, but we rarely have structured evidence of how. Synthetic data from calibrated simulations can begin to fill that gap.

The second direction is generative AI. The same datasets can train models that learn the relationship between policy inputs and spatial outcomes - making it possible to generate and evaluate design alternatives based on learned patterns, not isolated scenarios. This is not speculative. It is already the direction that tools like Autodesk Forma, Delve (https://www.bluelabellabs.com/work/delve/), and emerging urban foundation models are heading.

The specific role urban designers can play, and why it matters

This is where I think the profession has an underappreciated opportunity, especially in India.

In machine learning, there is a practice called Reinforcement Learning from Human Feedback (RLHF). In language models, human evaluators assess outputs and signal what is better or worse, and the model learns from that signal. The quality of the system depends directly on the quality and consistency of human judgement brought to that evaluation.

Urban designers are, in a very precise sense, positioned to perform this role for spatial AI systems.

Consider what experienced urban designers actually carry: the ability to read a street section and sense whether it will feel safe at night, to look at a land use mix and anticipate conflict or vitality, to recognise when a setback regulation is producing dead frontage rather than amenity. This is not vague intuition. It is pattern recognition built from years of navigating the gap between what regulations say and what cities actually do.

In the Indian context, that gap is particularly wide. Our cities carry layers of informal adaptation - setbacks converted to shops, FARs exceeded and regularised, roads that exist on master plans but not on ground. A purely algorithmic system trained on formal data will systematically misread this. Human designers who understand how Indian cities actually function (not just how they are planned) are essential correctives.

Concretely, this positions urban designers to contribute at two stages of emerging AI workflows:

At the pre-training stage: defining what variables matter, how spatial outcomes should be measured, and which relationships the model should learn. This is not a technical task — it is a design task. Deciding that pedestrian permeability matters more than plot coverage in a given context is a judgement call, and it shapes everything downstream.

At the inference stage: evaluating generated alternatives - flagging when an algorithmically optimal layout is socially implausible, when a density configuration ignores informal use patterns, or when a policy scenario is technically feasible but contextually absurd. In India especially, this requires someone who understands the distance between what a model can optimise and what a city can absorb.

A reframing worth considering

The anxiety in design professions about automation often frames the question as: will machines replace designers? That framing is probably not the useful one.

The more interesting question is: whose judgement will the machines learn from?

If urban designers do not engage with simulation, synthetic data, and AI workflows, those systems will be built anyway - trained on whatever data is available, evaluated by whoever is willing to do the evaluation. The values and spatial intelligence embedded in those systems will reflect that.

Developing fluency in agent-based modelling and structured data thinking is not, then, simply a technical upgrade. It is a way of ensuring that the accumulated spatial knowledge of the profession - including the hard-won, India-specific knowledge of how cities here actually work - remains part of how future design systems are built and used.

That seems worth doing deliberately, rather than leaving to chance.