AI-Based Simulation: Simulating Real Human Behaviour

A population-weighted, multi-agent simulation that recreates India's diversity using thousands of AI personas. Each persona is grounded in real census data on income, region, literacy and lifestyle, allowing them to debate, react and predict how any product or policy would actually land on the ground, all before a single rupee is spent on real-world testing.

It blends individual judgment with peer-influenced group dynamics to produce statistically calibrated adoption scores, segment-level insights and a strategic synthesis, all before a single rupee is spent on real-world testing.

Week one was about making Claude Premium do what we needed. Trust us, we tested it to its bounds. It failed, but never once told us to look for an alternative way. Legends say our prompt was wrong.

We ran into token limits, dead ends and a series of interactive dashboards that never became what we needed them to be. Every time we thought we were close, something else broke. We kept patching, rebuilding and trying again, staring at the possibility that all this effort might lead nowhere.

By the end of two weeks, we were exactly where we started: nowhere.

Armed mostly with AI, our non-tech team tried its hand at building the thing directly. We'd be doing a lot from first principles, because learning something and simultaneously building it forces you to question most of what you assume.

And that changes everything.

The plan, as pitched to ourselves over coffee: ask AI to write Python code, paste it into VS Code / Terminal / Python (whatever it's called; we're not tech bros here), watch magic happen, send in our sales wizards, and collect crazy profit.

The plan, as it spanned out: we found it hard to install Python or VS Code, running the code was harder, and honestly we're still not sure we installed all the dependencies.

After enough trial, error and mercy from people who actually knew what they were doing, we abandoned Terminal and VS Code and moved everything to Colab. The first attempt failed (that was OpenAI's code); it later worked with Claude's code on Google Colab.

The first output took 30 good minutes. We stayed glued to our screens, with every jumping line the code generated itself, like watching Picasso's strokes live.

Week 3 was about making vibe-coding work. We just wanted results and would give whatever it took, except learning Python properly, which wasn't possible in our project cycle. Turns out you don't always get to choose which parts of the process you're allowed to skip.

By the end of the week we had something that worked, and we were genuinely blown away by the outputs. Then the big legends questioned our base. So we argued with census data, fed it to our LLM, and ran the model with that context.

Still, our gut said it might not be enough. That a problem remained unsolved. That problem was waiting for us in week 4.