01
A scattered working day
A founder's day is spread across inboxes, dashboards, documents, calendars and chats. Talking to a chatbot about that work is not the same as getting it done.
Home / Playground / Watson
Own build: personal AI system
An assistant that works by voice.
A personal AI assistant that works by voice, with no hands, operating real tools the way a human operator would, and stopping at every security wall.
Working productPersonal system
The voice loop Watson runs on
SpeakGive the task by voice, no hands
Watson plansSpecialist agents take the parts
Real tools, real accountsWorks in the browser like a human operator
Safety stopsHalts at passwords, codes and security walls
Reports backEvery action logged and reviewable
Watson runs locally on my own machine as a private service with its own master dashboard. It remembers preferences, people and projects across sessions, and every action it takes is bounded by standing rules I set.
There is no public demo, by design: Watson operates private accounts and personal data.
Private by design.
A permission engine and a full audit log mean nothing happens outside the rules, and everything that happens can be reviewed afterwards.
The problem
A chatbot does not remove work.
02
Tools need an operator
The assistant has to operate the tools itself: browse, draft, file, track, and report what it did, not just describe what could be done.
03
Power without brakes is risk
An agent inside real accounts must remember standing rules, ask where it should ask, and stop dead at passwords, one-time codes and security checks.
What it does
What it does.
Voice-first controlSpeak a task and Watson carries it out. The full loop runs locally, with cloud voice as an option when a key is connected.
Works like a human operatorOperates inside the real browser and real accounts, with specialist agents coordinated under one assistant and one master dashboard.
Long-term memoryPreferences, people, projects and standing rules are remembered across sessions.
Permission engine and audit logEvery action is bounded by rules set in advance and can be reviewed afterwards.
Hard safety stopsPasswords, one-time codes and account security checks are walls Watson never crosses on its own.
How it works
Speak. Done. Reported.
STEP 01
Give the order
A task arrives by voice or message and is matched against standing rules and permissions.
STEP 02
Agents execute
Specialist agents carry out the parts: research, drafting, browser work, tracking.
STEP 03
Walls are respected
Any password, code or security checkpoint stops the work and brings it back to me.
STEP 04
Report and remember
Watson reports what was done, logs it for audit, and keeps what it learned in memory.
My layer
What I built.
Watson is an original build, designed and engineered by me end to end: the agent architecture, the voice loop, the memory system, the permission engine, the dashboard, and the service that runs it all.
It is refined through a continuous loop I run personally: test, analyse, fix, retest, with large automated test suites passing across its release line. The shipped product line is the v1.11 series, in daily use. A next-generation engine, combining two agent engines inside the same product design, is in development alongside it.
Watson is a personal system, built for my own work. It is not offered for sale.
Original build by Simanto Mitro.
Shipped and in daily use.
The v1.11 series runs locally on my own machine as a private service, handling real work every day. The next-generation engine is in active development.
Building something in health?
If this kind of system would help your clinic, organisation or team, the conversation starts with one cup of coffee.