Wafflestack

We build for the moment being wrong costs more than being silent.

Most AI products are judged on how much they can say. Ours are judged on knowing where to stop.

Where it started

Our first product was Mendly, a home repair assistant for Indian households. Building it taught us something uncomfortable. An AI that confidently tells a homeowner what to do with a geyser is dangerous in a way that a cautious one is not.

The failure mode was never that the model didn’t know. It was that it sounded as though it did.

Everything we have built since starts from that lesson. Arc Health shows a doctor seven years of creatinine values and does not tell them what they mean. The Apexvet formulary shows a veterinarian the published dose and does not suggest one.

The failure was never ignorance. It was confidence.

Three rules, and what each one costs us

A principle that costs nothing is a slogan. Each of these makes our products less impressive in some way, and we keep them anyway.

  1. We report. We don’t interpret.

    Our tools surface what the record says; the professional decides what it means. The cost: a tool that will not offer a diagnosis demos worse than one that will. We would rather lose that demo than have a clinician lean on a guess.

  2. Sensitive data stays on the device where it can.

    In Arc Health, patient reports are read inside the doctor’s own browser and are never uploaded. The cost: no cross-device sync of patient history, and harder engineering. That is the trade we chose.

  3. Our limits are governed controls, not preferences.

    What our systems may do, what data they touch and how their output is checked are managed under an AI management system certified to ISO/IEC 42001. The cost: process, documentation and an annual external audit. It is what turns “we’re careful” into something you can verify.

A tool that says “I don’t know” is less impressive in a demo and more useful on a Tuesday afternoon. We optimise for Tuesday.