In developmentMy own product · TresuraMaya by Tresura
An AI work assistant that remembers context, handles tasks and reminders, and only acts with your approval.
Scripted demo · no AI is called · answers and rules come from Maya’s own code and facts
Overview
Maya is the AI assistant for the whole Tresura workspace. You can talk to it or type. It keeps notes, tasks and reminders, drafts reports, emails and documents, and works inside Ledger with a signed sign-in handoff.
It is in development. I use it every day, a few testers have their own separate accounts, and Ledger staff reach it from inside Ledger.
My role
I designed Maya and lead its build. That covers the assistant, its tools, the mobile apps, and the monitoring setup that watches the other Tresura products.
I also designed a staged incident workflow. When a product alert fires, one AI agent (Muse) investigates, reports and tries a fix on staging. A second agent (Codex) checks that work or takes over. Committing and deploying to production always waits for my approval.
How Maya works
Every message goes through the same six steps. Click a step to see what happens there.
From the web app, the phone app, Telegram, or inside Ledger. Typed or spoken; speech is transcribed by Groq Whisper.
Maya works out who this is and what they may see. Inside Ledger, each person only reaches their own data, and Ledger decides what their role permits.
Before the model answers, Maya pulls in grounding: up to four verified facts, the person’s saved memories, and the real text of any file they mean.
The reply is written by Groq (gpt-oss-120b). If Groq is busy, it falls back to OpenRouter. Claude is available as a manual, owner-only choice.
If the reply wants to do something, such as sending an email, the server creates a proposal and waits. Risky actions need a two-step confirmation.
Text, plus cards and files (PDF, DOCX, XLSX). It can be read aloud with Cartesia text-to-speech.
Grounding, not guessing
Maya looks things up before it answers (the idea behind RAG). I chose the lookup method on purpose:
Facts
- A small, verified set of facts about the products and company
- Matched by weighted keywords; names like “Ledger” or “Relay” count most
- No embeddings and no model call on this step, so it’s fast and the tests can predict exactly which fact wins
- At most four facts per message, marked as overriding what the model thinks it knows
Files
- Uploads, including from Telegram, go into one File Vault with their extracted text
- Answers come from the file’s real content
- If the text isn’t available, Maya says so instead of inventing a summary
Memory
- Secrets such as passwords are never stored, even if you ask
- Sensitive topics are saved only when you clearly ask
- Anything Maya merely noticed is a suggestion until you say yes
- Memory can be switched off; nothing is deleted when you do
Actions
- Only your own message can authorise an action
- Text inside a file, a web page or a memory is treated as data, never as an instruction
- Risk levels are set on the server, so the model can’t change them
- Knowing who you are isn’t the same as being allowed: Ledger checks your role every time
When a product breaks
Maya also watches the other Tresura products. I designed the incident loop so AI agents do the legwork and I keep the final say.
- DetectA product monitor checks Ledger, Maya and Boost, and sends a phone alert when one goes down.
- InvestigateMuse, an AI agent, investigates, writes up its findings and tries a fix on staging.
- VerifyCodex, a second agent, checks Muse’s work, or takes over if it’s stuck.
- ApproveI review it. Nothing is committed or deployed to production without my approval.
What it does
Assistant
- Chat by text or voice, with memory across conversations
- Tasks, notes, reminders (recurring, Malaysia time), files, projects
- Drafts reports, emails, documents and leave requests; exports PDF, DOCX, XLSX
Safe actions
- Sending email or Telegram needs a two-step confirmation
- Every high-risk action waits for approval
- Inside Ledger, each user only sees their own data
Voice
- Speech-to-text: Groq Whisper
- Text-to-speech: Cartesia
- Push-to-talk on mobile
Operations
- Health dashboard and phone alerts when a product goes down
- Incident reports reviewed from the phone
- Android app updates itself in-app
How it’s built
By the numbers
Timeline
- 15 Jun 2026First commit.
- Jul 2026Android app with push-to-talk voice.
- 14 Sep 2026New text-to-speech voice live in production.
- 25 Sep 2026iOS 1.6.0 redesign uploaded to TestFlight.
Where it stands
Not publicly launched. WhatsApp and Gmail, Drive and calendar connectors are planned, not built. Actions Maya can take inside Ledger are built but switched off.