I started in design, moved into product, and kept following the work until I was building the systems too. Today I’m the founder and sole engineer of Vaulth AI.
Explore the work → What I'm building now ↓ Read what I think →
These are the parts I still think about: the choice I made, what I gave up, and why I’d make it again.
Vaulth AI · architecture
Seven caregiver scopes enforced across 62 API endpoints, every external view token-gated and revocable, family circles that share a trend and never the raw number.
I said no to: taking custody in exchange for access. It would have been faster to build and easier to fund, but it would have put the company in control.
Read the decision →Arth · architecture
Bank SMS, PDF statements, and CSV exports converge on one on-device pipeline: deterministic rules for the common case, a small on-device model only for the uncertain tail.
I said no to: using the model for everything. That would have made the app slower, more expensive to run, and harder to trust.
Read the decision →Cricket Manager · judgment
A 0.98 correlation on the ratings model was the most flattering number the project ever produced. I read it as evidence of a problem and went looking for the leak.
I said no to: using it as the headline result. The number was technically true and still gave the wrong impression.
Read the decision →A family's medical records should answer to the family.
In India those records are scattered further than the software assumes — hospital silos, WhatsApp threads, prescription drawers, lab portals — and every existing answer has an owner problem. Portals hold records hostage to one provider; aggregators take custody in exchange for access. Vaulth's answer is that the family is the owner, and that consent, not storage, is the product. I built the consent architecture first and hung everything else off it.
The case study — ten decisions and what each cost →
vaulthai.health — the platform itself
Can a finance app understand your money without ever being told what it is? Bank SMS, PDF statements and CSV exports parsed on-device by a deterministic-first cascade, with a small on-device language model for the long tail.
0 bytes transmittedphase 4 of 5 — unreleased
How do you make a simulated cricketer behave like the real one — and how would you know if you had failed? A ball-by-ball engine calibrated on 2.07 million real T20 deliveries, player ratings adapted from the Earth Similarity Index, and a live event-sourced auction.
2.07M deliveries modelled 0.98 correlation — rejected as headline
Ten projects across product platforms, independent systems, and research-led builds →
I have a hunch that source code is becoming another layer most people won’t work in directly, much as assembly did before it.
If that’s even partly right, the interesting work moves toward choosing the problem, judging the product, taste, systems thinking, understanding people, and distribution. The essay is me trying to make that case honestly, including the evidence that makes me less certain.
Software is learning to write itself. Someone still has to know what's worth writing.
If you’re building something interesting, or disagree with something here, I’d be glad to hear from you.
Let’s build something
worth caring about.