Live product · 2026
Hailera
A “Google for AI agents”: structured access to live business data and knowledge through MCP. Startup participating in the EWOR program.
- Role
- Design Engineer / Sole Technical Contributor
- Year
- 2026
- Stack
- Next.js, TypeScript, Supabase, Python, AWS, MCP, Figma
An agent a business owner can actually trust
Hailera gives a business its own agent. It answers questions from ChatGPT, Claude and the other assistants people now ask first, books appointments, and reports back to the owner. I was the sole technical contributor, so I designed it and built it: the interface, the design system, the app, and the agent infrastructure behind it.
The hard part was never the AI. It was making an owner comfortable letting something answer for them while they are busy, or asleep.
How it should feel
Calm, a little warm, and never mysterious. Every number says where it came from, every answer shows what it was based on, and the owner always knows what the agent did while they were away.
I wanted it to feel like a capable colleague giving you a short update, not a control panel asking for attention.
Up close
Where the questions come from
An owner opening the dashboard wants to know one thing: is the agent being asked, and by whom. So the biggest number is the questions count, 2,764 over twelve months, with the growth chip and the crawler visits beside it in small type. The chart is one stacked bar per month, with each assistant a different colour and named with its own count under the chart. Hovering a month opens a small card that lists every assistant with its count and share, so the exact figures are one gesture away and not crowded on the surface. The four bookings and questions tiles above it use the same pattern, a number, a change chip and the sources underneath.
Ready in minutes, not a setup project
A new owner is busy and wants the agent working without a setup project, so onboarding starts from what already exists: the business’s own website. A four-step list on the left shows where you are, with a green tick for finished steps and a black circle for the current one. The centre card says Your agent is ready and shows what was found, with each detail labelled by the page it came from, such as Contact page or Header, so the owner can check facts quickly instead of retyping them. Knowledge chips show the documents written, and one is still writing, which shows the work is live. A single black button, Check your details, is the only next step.
Test before you trust
The owner’s real worry is what the agent will say to a customer, so I built a place to find out in private. They ask as a customer would, and each answer starts with a small line, Answered from Prices and Roof inspection, so the source is visible before the text. The second question shows the more important case: when the agent does not know, it says so, gives the business’s phone number and does not guess. A card then explains the gap in plain words, Your agent didn’t know this, and offers one blue action, Add an answer. A footer reading Customer view, nothing here is public removes the fear of testing in front of anyone.
An agent that shows you around
A product tour is usually a row of tooltips that people dismiss, so I let the agent do the introducing instead. Its message is written in the first person and says what it does for the business: it answers people about the business on ChatGPT, Claude and the website, and it offers a one-minute look around. The reply is one blue button, Show me, and the agent’s own face sits next to the message, so it is clear who is speaking. The dashboard stays behind the message, which keeps the real product in view while the owner is being guided. The sidebar groups, such as Reach and its items, give the agent real places to point to.
Updates that respect your day
An owner does not want a feed, so the notifications panel separates what needs a decision from what is only news. A request, Waiting for your approval, comes first with the proposed change, Turn on Sunday bookings, its time window, a note that it expires in 23 hours, and two actions: a filled blue Approve and a quieter Review. Below it, the news item, Anna Kowalski booked a roof inspection, carries the date in a small orange chip and a single View booking link. A blue dot marks unread items, and the Unread count in the tab tells you how many to expect. Everything is written as a short sentence in plain words.
Looking back
Hailera became a working product that I designed and built end to end as the sole technical contributor. What I take forward: trust in an agent is designed in small moments, like a source under an answer, a note on a booking, or a clear “nothing needs you”. The test screen matters most here: admitting what the agent does not know is what makes the rest of its answers believable.