KT Gilcrease
Platform · Case Study
Full case study

Yare & TasteBud(s)

Platform · Case Study
← Back to projects

The idea

Most AI apps give you one assistant that tries to do everything, and it shows in generic advice, no real personality, no sense that it actually specializes in anything. I wanted the opposite: a curated circle of AI companions, each one genuinely expert in one narrow thing, that you actively choose, promote, bench, and swap the way you'd manage a real relationship. Yare is the platform I built so that idea could power any domain. TasteBud(s) is its first product, the idea applied to food, with a roster of culinary "Buds" like a breakfast specialist, a fermentation expert, a foraging guide.

What it does

TasteBud organizes your AI companions into tiers: a small, ranked "A-List" you actually talk to, a Watch list of prospects, a Bench for companions you're not using right now, and an Archive for ones you've set aside. Every Bud keeps its full memory of you no matter which list it's in. Archive one for six months, bring it back, and it remembers exactly where you left off. Your A-List is capped by subscription tier (3, 7, or 12 slots), and when it's full, swapping in a new Bud means explicitly choosing who steps down. Nothing is ever silently evicted.

How I built it

Yare is the engine underneath: a shared platform for relationship modeling, memory, and companion behavior, with each app (TasteBud for food, an early second app called Draft for writing) as a themed skin on top of the same core. Every Bud is authored against a consistent behavioral template (personality, tone, domain boundaries, safety rules), so a companion can't wander off its lane no matter how the conversation goes. That was the whole point: I'd noticed every hyper-focused chatbot I tried eventually drifted, and I got obsessive about engineering a structure that actually holds the line.

Development itself became a case study in something I care a lot about: running an AI team well. I split the remaining build into two parallel workstreams, one AI agent on state and data logic and another on UI and screens, with a shared interface contract between them so neither could break the other, and every change reviewed before it merged. Nothing shipped without that review.

What I was exploring

Beyond the app itself, I wanted to answer questions most people building on AI models don't bother asking until it's expensive to find out: What does this actually cost to run at scale? What happens to a companion's personality when the model underneath it changes? I modeled the hosting economics in detail and found something counterintuitive. The "obvious" choice of renting a dedicated GPU actually loses money below roughly a thousand paying users, while pay-per-use hosting stays under a cent per user even for heavy usage. I also designed for a reality I don't think gets taken seriously enough: AI models change constantly, and swapping one under an unchanged personality can quietly break behavior nobody's watching for. So I built a way to test a companion's behavior against a fixed standard every time the model underneath it changes, treating that instability as a permanent condition to design around, not a one-time migration.

My role

Concept, product design, platform architecture, cost modeling, and AI team direction, end to end, with two AI coding agents executing against specs and contracts I set and reviewed.

What's next

TasteBud is still pre-launch. The state layer and UI are being built out now, live AI conversation isn't wired in yet, and the paywall is UI-only so far. Yare, the platform underneath, is built to support more apps than just this one.