Design and data, treated as one discipline
I'm an AI Product Manager who works with founders and teams as a product consultant. I care about the half-second before feedback, the sentence a user has to defend to their boss, and the feature we choose not to build.
My work spans on-device consumer apps where latency and privacy are the product, and enterprise NLP that has to ship into regulated, audited environments. The common thread is trust: capability is rarely the bottleneck, but whether a person will rely on the system twice almost always is.
As a consultant I tend to come in on the hard middle: framing the eval before the feature, sequencing a roadmap as a trust-building order rather than a capability list, and being honest, in writing, about what I'd refuse to build. The case studies and teardowns here are the clearest picture of how I actually operate.
- Discipline
- AI Product Management
- Mode
- Product consulting
- Focus
- Trust, evals, on-device
- Based
- India · remote-friendly
Six beliefs I manage AI products by
Design the eval before the feature
The scariest thing about a probabilistic product is that "it feels better" is not evidence. I freeze a representative test set and a scoring rubric first, so every model, prompt, and design change is judged against something that does not move. What you choose to score is what you actually ship.
Trust is the real roadmap
For AI products, capability is rarely the bottleneck; trust is. Latency, legibility, reversibility, citations, and honesty about uncertainty are not polish; they are the features that decide whether anyone uses the thing twice. I sequence work as a trust-building order, not a capability list.
Subtraction is a product decision
The strongest calls I have made were about what to refuse to build: the cloud path, the extra feature, the engagement hook. Minimalism on the surface is earned by obsessive depth underneath: one loop done completely beats ten done halfway.
Design and data are one discipline
I do not hand a spec to design and a metric to analytics. The half-second before feedback, the phrasing of an uncertain claim, and the number that proves it are the same problem viewed from different sides. The interface is where the model becomes a product.
Start from the person who gets blamed
The real buyer is often not the user of the feature but the person accountable if it is wrong: the parent, the risk officer, the on-call. Designing for the signer, not just the builder, is how AI products cross from impressive demo to something an organization will stand behind.
Context beats model choice
Two teams on the same model ship wildly different products based on what they put in the context window. I treat retrieval, grounding, and the shape of what the model sees as first-class product surface, usually more decisive than which model is underneath.