Venkat Sadras
Miya, drawn as a smiling badger in a dark kurta and a gold scarf, holding a glass of chai, with biscuits beside it

My personal AI agent · leads the team

Miya

My personal AI agent. One place to ask about my training, my evenings, our weekends and where we eat, with a team of specialists behind it.

  • Lead
  • Live since Jun 8, 2026

What I ask Miya

One place to ask. A specialist answers.

I don't pick an agent. I just ask Miya, and the ask goes to whoever owns it. Here's who answers what, and how.

  • Kobe answers

    “Where am I this week?”

    Answered in code from the watch data. No model, same answer every time.

  • Fraser answers

    “Design me a session for tomorrow.”

    Written by a model, then checked against a 10-point contract in code before I see it.

  • Huberman answers

    “Give me a cooldown.”

    Built from a 25-drill library for the day I actually had. It also shows up on its own every evening.

  • Genie answers

    “What's on this weekend?”

    Local events from 121 sources, planned around what we've already decided.

  • Bourdain answers

    “Where do we eat?”

    On its own bot for now, being tested on the harness. It joins Miya when it passes.

  • Miya answers

    Anything else

    No rule claims it, so Miya's model takes it, asks the specialists for facts, and answers in one voice.

Checks in the path

It checks before it speaks

A validator compares each reply with what's actually true, with seven checks, and rewrites it before it sends.

One voice

Many agents, one voice

Specialists do the work; Miya answers. Exact numbers from code go out word for word.

A record

Every decision written down

Which agent, which rule, how long, what it cost: 7,408 decisions logged since May 8.

The problem

I didn't want six bots in my pocket. I wanted one place to ask anything, and an assistant that wouldn't confidently tell me something false about my own life.

Why I have it

Miya is the front desk and the editor. It sends each message to the specialist who knows the answer, and checks that answer against what's actually true before it reaches me.

What it does

  • Routes every message: rules in code answer the asks they recognize, and Miya's model handles everything else.
  • Pulls facts from the specialists and answers in one voice.
  • Sends a morning brief, and one line when something in the system is broken. Silence otherwise.

Where it helped

  • After the rebuild, I replayed 126 real messages: internal-voice leaks went to 0%, and contradictions were caught and corrected in about 5% of replies.
  • 7,408 decisions logged since May 8, each with the agent, the rule, the latency and the cost.
  • On its first day on the new platform it surfaced a bug that flipped "ahead" and "behind". It's now pinned by a test.

How it works

One message, step by step

Each step lights up in turn. The names are the real components in the code.

  1. TelegramI send a message.
  2. Miya's runnerGives it a trace id, so every step can be followed.
  3. The ladder16 rules in code. The first to claim the message answers it, with no model.
  4. Miya's modelTakes what no rule claims, and asks the specialists for facts.
  5. One voiceThe reply is put in one voice, unless it's an exact number from code.
  6. Validator7 checks against what's true. Rewrites the reply before it sends.
  7. Back to meEvery step lands in the decision ledger.

Not built yet

  • Routing each ask by what every agent declares it can do (a manifest) is planned, not built.
  • The plain-language intent layer is built but switched off.
  • Memory consolidation hasn't been scheduled since May.
Read its code →
← Bourdain All agents Kobe →