K.A.R.A is our own AI. It listens, answers out loud and runs errands for the team, and during the Bangkok flood it wrote the city's briefing every three hours from nothing but real data.
During the Bangkok flood, K.A.R.A wrote the city's daily briefing.
Every three hours, K.A.R.A reads everything the Bangkok Situational Map has published, the sensors, gauges, help requests and the news it has checked, and writes one calm summary for everyone, in Thai and English: where it's serious now, what's changing, a few gentle suggestions and what to watch next. It opens once a day on the map, and lives in its own tab.
officialDeepest street sensors, and whether they're still reporting
officialHeaviest rain in 24 hours
officialCanals over the bank, rising and falling
officialEach district now vs three hours ago
officialDams near capacity, roads to avoid
officialHelp requests in the data
officialAirport weather and aircraft movements
news·AIOfficial alerts and closures found in the news
news·AIAirport operations reported in the news
modelNext high tide at the river mouth
The rules it writes by
01Only what the map already knows. No numbers of its own.
02Official data first; news only fills gaps, and is said as ‘reports say’.
03If a source has gone quiet, say so. Never claim ‘steady’ without data.
04Name places, don't count them. The app shows its own numbers.
05Calm, kind, practical. Invite, never order.
06Thai and English, every time.
On the map, K.A.R.A's words come from a Google Gemini model, held to these rules and to the map's own data; our studio K.A.R.A runs on our own machine. A community tool, not an official warning.
K.A.R.A at work
Thirteen feeds in, one calm map out.
Every source passing through K.A.R.A on the Bangkok flood map: read, weighed, aggregated, purged, and placed. Drag to turn it.
Government dataReports from peopleNews read by AIEvery dot is a real report or reading, falling from top to bottom: through each source's door, collected, read and weighed by K.A.R.A, anything unreliable purged, and what's left placed as a pin on the map of Bangkok.
33%of enterprise software expected to include AI agents by 2028, up from under 1% in 2024Gartner, 2024 ↗
Industry research, linked to its source. Our own numbers below come from the Bangkok flood map.
What it opens up
Five ways an assistant like K.A.R.A helps everyone.
01
Development
Assistants that read the codebase, draft, test and explain, so a small team ships like a bigger one and spends its time on the hard parts.
We build with AI in the loop every day: the flood map went from idea to a city-wide live service in days.
02
Analytics
Many feeds read at once, weighed by how trustworthy they are, and turned into a plain summary anyone can act on, with the sources kept.
K.A.R.A read 13 feeds and wrote Bangkok's flood briefing every three hours, from real data only.
03
AI as a service
Your own assistant, trained on your documents and rules, offered to your customers or staff through an app, a site or a chat, and kept honest.
On the flood map, K.A.R.A runs as a service: every three hours it writes the briefing the app shows to the whole city.
04
Software for agents
Agents that don't just answer but act: check a system, fetch a file, fill a form, hand off to a person when it matters. Built with guard rails from day one.
K.A.R.A runs errands for our team: it listens, answers out loud and does the small jobs in between.
05
Business automation
The repeat work, reports, checks, sorting and replies, done on a schedule and logged, so people keep the judgement calls.
On the map, collection runs every five minutes and moderation purges stale or unsourced points before anyone sees them.
We keep our own model. It runs on our machines, it has read our work, and your material never leaves the building.
Knowledge Augmentation Realtime Assistant. It drafts, checks, searches and summarises while the team designs, and answers out of our own record — past builds, specs and the decisions behind them — rather than whatever the open internet last said. It proposes and a person decides, which is the whole of our position on it. AIEngram, the instrument drawn here, is how we look inside it.
Our attitude towards AI
A tool we delegate to, not a replacement
Experiences that adapt
More intelligent, more personal, different for each visitor.
Live sooner
Shorter build, faster deployment, the date met with room to spare.
Audited harder
Every screen, flow and edge case checked, not just the happy path.
More revisions
Versions tested against each other instead of one idea defended.
Your material stays here
K.A.R.A, our own model, runs on our machines. Nothing goes out to train anything.
At no extra cost
The rate is the rate. The time saved goes into the work, not the invoice.
With the model, and without
Same team, same brief, counted both ways
Four measures, each taken on our own work. The red bar is the job run with K.A.R.A in the loop; the grey bar under it is the same job run without.
42%shorter
Time to a working build
With K.A.R.A3.5 weeks
Without K.A.R.A6 weeks
56%fewer
Defects that reach you
With K.A.R.A4 per project
Without K.A.R.A9 per project
2.4×more
Versions tried
With K.A.R.A12 versions
Without K.A.R.A5 versions
5 daysto a first draft
First draft in your hands
With K.A.R.A5 days · by Friday
Without K.A.R.A3 weeks
How they are counted. Time to a working build is kickoff to a build you can hold and try, not a demo video. Defects are issues found after handover, per project, whatever their size. Versions are options built and tested before sign-off, not slides of options. The first draft is something to react to by the end of the first week, not a document about it.
AI in the work
Two builds that read the person in front of them
Both from the same three-week job for Thailand Post. One ran all weekend. The other was finished and never switched on.
Thailand Post · Talat Noi · October 2025
Animal Reaction: a booth that reads the face
What it does. A visitor stands in front of the booth and pulls a face. The screen answers in real time with the illustrated animal whose expression matches theirs.
Where the model sits. Facial expression read frame by frame from the live camera, mapped to the character set the builders drew for the activation.
Why it works. No instructions, no buttons, nothing to read. A child understands it in a second and the queue explains itself.
Thailand Post · Talat Noi · October 2025 · built, not used
Letter From My Future Self
What it does. A visitor answers a handful of questions about where they are now. A letter comes back, written to them by themselves, years from now.
Where the model sits. The letter is written on the spot out of what they actually said, in their own words and their own register, rather than assembled from templates.
Why it works. The person is the content. Nobody throws away a letter addressed to them in their own voice.
