On-device by design
In our MVP, every agent runs on the phone in the user's hand: language model, knowledge search, and translation. Recovery data is processed where it lives.
AI software for wellness professionals and venue operators
Recovery AI that learns from everyone, while your data stays yours.
That is the end state we design for: AI that gets better across people, professionals, and venues, while each person's data stays on their own device. We start with the people who deliver recovery. Wellness professionals and venue operators get AI agents that run on their own phones, and a human approves every step.
MVP: three on-device agents running end to end on real phones. Piloting in Taiwan with our first partner professional association.
Where we're going
We started from the end.
Most products bolt privacy and shared learning on at the end. We made them the starting point, then worked backwards to what we can ship today.
Personal models that learn privately
Small personal adapters that learn how each person recovers, trained for that person and kept on their device.
Learning across the network
With federated learning, venues and associations improve a shared model by exchanging model updates, never raw personal data.
The goal is better recovery for more people, everywhere, without anyone having to give up their data. That is the deep tech we think is worth building, and it is why Balance Project exists.
What we're building
Three agents, one network.
Recovery falls apart in the gaps between the person, the professional, and the venue. Each one gets an agent, and BOP, our open protocol, connects them.
Pro AI
Serve more clients.
Drafts notes, follow-ups, and plans from the association's own materials. The pro reviews and approves every message.
Personal agent
Say what you need.
Turns a sentence into a clear request and shares only what you approve: a summary for your professional, and just the service, time, and budget for the venue.
Venue agent
Fill sessions, stay in control.
Answers requests in batches, confirms bookings, and keeps prices and capacity under the operator's control.
MVP
Working today, on a 2021 mid-range phone.
All three agents run end to end on real phones and reach each other directly, with no server in the middle. These screens come from a Samsung Galaxy A32 5G running on its CPU alone. The agents reach each other over encrypted peer-to-peer connections. The venue and knowledge base are demo content.
Screens shown in English; the app also runs in Traditional Chinese.
How it grows
Each association makes the next one better.
Association
A professional association brings its course materials and expert judgment.
Tuned Pro AI
We tune Pro AI on the association's materials and expert labels.
Tested before launch
Each version is tested before it reaches the association's pros.
Pros serve clients
Members use Pro AI with their own clients and approve every message.
Consented feedback
Feedback shared with consent improves the next version, and then the next association.
How we build
Open protocol, on-device AI, humans in the loop.
BOP, the Balance Open Protocol, is an open service-contract profile on top of existing agent and health standards. It is not a new standard, and it is free to adopt. BOP describes how the agents for the person, the pro, and the venue agree on a service, record consent, and confirm that it was delivered.
Open core, paid network. The protocol is open; we earn by operating the network and the services that run on it.
Built on
- A2A
- MCP
- AP2
- UCP
- OpenActive
- FHIR
An agent on every phone
The person, the professional, and the venue each run their agent on their own phone, and the agents reach each other over encrypted peer-to-peer connections.
Claude in our workflow
We build with Claude and Claude Code for research, specifications, and code. We are also developing a BOP reference MCP server that uses Claude as the client agent in end-to-end tests.
Reviewed before release
Before any Pro AI version reaches professionals, it is checked against evaluation sets for out-of-scope advice and missed referrals.
Principles
Built for trust from the first session.
The pro stays in the loop
AI drafts; the professional decides. Nothing reaches a client without the pro's approval.
Consent travels with the data
Your recovery context moves between providers only with your consent, item by item.
Open by default
BOP is open and free to adopt. It builds on standards the industry already uses instead of inventing a new one.
Work with us
Who we partner with
Professional associations
Bring your materials and expert judgment. We tune Pro AI for your members, test it, and launch it with you.
Venue operators
Recovery studios, saunas, and gyms: requests, bookings, and proof of completion handled by an agent you control.
Enterprise wellbeing teams
Recovery programs for employees, priced on participation and completion.
Team
People who have lived every side of the network.
Eliza Felos
Founder
15 years of competitive swimming, including 3 as an NCAA Division I athlete and team captain. Leads product, service design, and international expansion, and is preparing a founder-run wellness venue in Taiwan as our first pilot site.
Anthony "Tony" Felos
CTO
Founded SurfMerchants, maker of the SASSIE market-research SaaS platform, acquired by LiveShopper in 2020; then CTO and CISO of the combined company. Now builds enterprise AI for global energy and banking clients.
CN Lin
Executive Advisor
15+ years in cross-border payments, partnerships, and APAC market entry.
Contact
Let's build recovery that holds together.
Professional associations, venue operators, enterprise teams, and investors: we'd like to hear from you.