On building Jimbo, our AI agent for the messages people actually send
Most "AI for business" lives behind a login nobody opens twice. We wanted the opposite: an agent that answers where customers already are — the thread they're already in.
Meet people in their inbox, not yours
A restaurant's customers don't want a portal. They want to ask "are you open tonight?" on WhatsApp and get an answer in seconds. So Jimbo's first job was never "be clever" — it was "be there." We started with WhatsApp, then Instagram, then SMS and Facebook, because that's the order real businesses asked for.
The hard part isn't the model. It's the seams between channels: identity, history, and tone all change shape when the same conversation moves from a DM to a text. Getting those seams invisible took longer than the reply logic did.
Boring reliability beats clever answers
We learned to optimise for the unglamorous things first — delivery, latency, knowing when not to answer and hand off to a human. An agent that's right 95% of the time but escalates the other 5% gracefully feels trustworthy. One that's right 99% of the time but fumbles the handoff does not.
What's next
Jimbo gets sharper as it learns a business's own context — its menu, its hours, its voice. That's where the Restroverse integration comes in, and where the next note will go.
More notes
The quiet work of software
Most of what we build is invisible by design. A note on why we treat that as the goal, not a compromise.
Handing Shramdan to the community that built it with us
A short essay on starting something you intend to give away — and the practical shape of doing so.