
A dashboard that shows teams what their AI agents cost and recommends how to spend less without the agents getting worse.
- Scope
- Product, design system, web, brand, pitch
- Timeline
- 2026
- Role
- Cofounder and CPO
- Team
- Mehek Mandal, Kyna Rochlani, Sania Gupta
Setup
Traeco is a dashboard that shows what AI agents cost and how to spend less
I was cofounder and CPO with Mehek Mandal, Kyna Rochlani and Sania Gupta, and as the only designer I designed and built the product, the design system and the website, and shipped all of it to production.
Showing teams what they could save was easy. Getting them to act meant proving their agents wouldn't get worse.
The dashboard home: savings available, then the top recommendations.
- 22% to 71%acted on a recommendation, in usability tests
- Audience ChoiceLavaLab closing summit
- shipped to production
Shipped
I designed and built the dashboard and shipped it to production
Drilling from a recommendation into one agent's six-week quality history.
- 01
Home
Savings, top recommendations and spend.
- 02
Recommendations
Cost, quality and confidence, with the option to test before committing.
- 03
Agent detail
Six weeks of quality history for each agent.
- 04
Quality budgets
Limits a team sets once on how far quality may slip, and the product holds every change to them. Set through the API today; the settings screen is next.
Problem
Teams knew agent costs would become a problem, but it was not urgent yet
We were 50+ calls into looking for a startup idea across 11 industries when our own AI agents started costing us more every week. We narrowed the rest of our research to teams using agents and kept hearing the same thing. Plenty of tools showed what agents had spent, but almost none we found told you what to change.
Fifty-plus calls across eleven industries, narrowing to agentic AI.
It isn't only startups. Uber exhausted its entire 2026 AI coding budget by April and now caps engineers at $1,500 a month per tool.
Decision 1
Because it was not urgent, I made setup two lines of code
Nobody would sit through a long setup.
Decision
Savings available sits at the top of the dashboard, so the first thing a team sees is money they can get back instead of a bill.
What I learned
When a problem isn't urgent yet, every setup step is a reason to say no.
import osfrom anthropic import Anthropicfrom traeco import init, wrap

Add two lines around an existing client and the first dashboard view fills in from the traces.
Decision 2 · Exploration
The early home screens stacked every number on one page
The early versions led with a sentence about last month's spend and three recommendation cards, first under the Slash name in orange, then as Traeco in green, with the spend breakdown and the at-a-glance numbers stacked below.


The early versions: Slash, then the first Traeco.
Decision 2 · Solution
I cut the home screen down to what people would pay for
I sorted everything from the calls into two lists, what makes the product usable and what makes it worth paying for.
Decision
Home shows savings, top recommendations and a spend breakdown, and everything else is one click deeper.
What I learned
The list of what people would pay for is always shorter, and it makes the better product.


Drag to compare my first iteration with the most recent one.
Decision 3
I showed quality next to cost so teams would trust a recommendation
Recommendations were what set us apart, and they also made people nervous; if a cheaper setup made an agent worse, that was on us. So we tracked quality on every agent and tested each change before it rolled out.
Decision
Each recommendation shows cost, quality and confidence side by side, and teams can try it on a small share of traffic before switching fully. Quality next to cost is what removed much of the stigma and the early hesitation from the first users and partners we worked with.
What I learned
A recommendation is only as useful as people's trust in it.
Open a recommendation to see what it saves, what it does to quality and how sure the judge is, then try it on part of the traffic.
In code
It was the first product I built in code and shipped myself
It was also my first time using Claude Code. Seeing real data in it showed me quickly what was feasible and what only worked in Figma, and every component had to hold up for one agent or hundreds.
It made me realise that the best designers know how to properly integrate AI into their workflow. Once I got comfortable, I was shipping ten times as much, exploring edge cases, changing and iterating faster than ever, and I had to teach myself how to scale, which I never really knew how to do before.
Around it

Brand
Near-black with a single green, so the only colour on screen is what matters.

Website
traeco.dev walks a buyer from the problem to pricing.
Pitch decks
One for presenting live and one that works on its own in an inbox.
What I learned
I learned that a product needs time spent on one idea
I enjoyed the earlier explorations and the industries they took us through, chip manufacturing for startups among them, but getting here sooner would have let me sit with the product and flesh it out properly.
Next time, I'd commit to one idea earlier and spend more time on it.