Welshare Health builds a health profile matchmaking protocol (HPMP) that allows AI-driven medical researchers to meet their ideal human participants: users who selectively contribute their health data to scientific research. Agents can execute queries to find correlations on the whole set of user data without sacrificing the privacy of individuals.
As Sam Altman noted in February 2025, AGI will accelerate scientific progress faster than today, potentially outpacing all other effects. This acceleration is mostly caused by the appearance and explosive development of AI agents - sophisticated systems with LLMs at their core that are able to do complex knowledge work, reason about tasks and split them into executable plans, talk to the outside world and grow memories over longer periods of time than one chat window.
From here, autonomous AI research agents actively drive scientific processes from hypothesis to onchain IP inside an ecosystem that also tries to protect user privacy and rewards participation - a future we commonly refer to as the "Transformation Epoch of DeSci" or the second version of the Decentralized Science movement and that will see the onboarding of large amounts of users into the space and the rise of agentic AI research.
The biology research process produces over a million new publications yearly. No human researcher can keep up with this volume of information. In contrast, agentic researchers are continuously running applications operated inside small labs or in the cloud that can react to environmental triggers and constantly collect and synthesize new information, regardless of its size. They accelerate the speed at which research papers are turned into validated IP and they become the foundation of an incentivized, permissionless, user-operated, decentralized ecosystem of a science process elevated by large language models and knowledge graphs.
Despite the powerful autonomy of AI agents, the trajectory from hypothesis to validated discovery can not be entirely automated. While agents excel at identifying correlations, generating hypotheses at scale, and interfacing with diverse external systems, they remain unable to directly observe the real-world implications of their innovations.
In the currently prevailing ecosystem, machine assisted research predominantly relies on pre-existing user data, or on synthetic derivatives. Although invaluable for detecting intricate statistical correlations and formulating novel hypotheses, these approaches fall short when validating the original assumptions that agentic researchers postulate; they neither can access data that hasn’t been disclosed to them nor can they predict how behavioural changes would affect the individual user’s condition.
To validate their ideas, agents would substantially benefit from adding a "patient in the loop" into their workflow instead - a genuine human element that they can ask to share measurable data readings or subjective observations that they cannot extrapolate from a pre-existing data set. In essence, agents will need to verify their assumptions directly with patients or human subjects relevant to their research space.
A patient-in-the-loop isn't necessarily a simple one-off interaction but can potentially grow into a sophisticated dialog-based process occurring over time to substantiate the real-world relevance and accuracy of the agent's discoveries or theoretical constructs. Imagine a type 2 diabetic who shows signs of remission and an agent that researches phenotypes that respond well to early time-restricted eating. The agent could get in direct contact with the individual, compare them with other patients in the remission group, unobtrusively ask for insights of their eating behaviours and find correlations between entities in the responder cluster. The collective findings might lead to a better foundation for health apps developing algorithms that help the individual monitor and understand their calorie intake schedules.
The critical challenge lies in enabling this human-AI interaction without compromising the equally important need for patient privacy and data control - a well known problem domain that’s addressed and made possible by state of the art privacy enhancing technologies like Nillion’s blind modules, Vana’s encrypted data storage or generic distributed computation protocols like Lit Protocol.
Welshare proposes a Health Profile Matching Protocol (HPMP) that allows research agents to query a vast amount of user profiles for certain conditions they've identified during their hypothesis generation process. The filtering and matchmaking logic runs inside an isolated but distributed trusted execution environment, e.g., a Nillion SecretVault on a dedicated cluster that's operated by trusted actors. Agents can run analytical aggregation queries on private information and use the result sets for their needs without ever disclosing any specific detail of a particular user. This already is a valuable feature to filter insights over large amounts of generally available user data.
However, a far more powerful feature of the HPMP is that it enables conversations between research agents and individuals who match a certain condition identifying them as a relevant target audience. Agents can communicate with users in accessible language, ask them to share additional data points, take individual pictures, observe and analyse behavioral patterns, or gather supplementary sensor data. The HPMP orchestrates the preservation and retrieval of conversational memory and contexts while guiding both human participants and agents through the process.
Research Agents will interact with the HPMP using dedicated Model Context Protocol (MCP) servers that translate requests into actionable operations within the HPMP interface. MCP is a foundational open standard introduced by Anthropic that enables AI agents, particularly LLMs, to securely connect with external data sources and tools.
Key features include:
Dynamic Access: Agents can query servers for real-time data rather than relying only on pre-trained knowledge.
Two-way Communication: Agents can both pull data and push updates through the server.
Standardization: One integration works for many different agent frameworks and clients.
Imagine a research agent tasked to develop new methods to detect biomarkers for early diagnosis of neurodegenerative diseases.
After the agent first queries the network for a promising correlation of conditions it files an aggregation query requests on HPMP to identify users who
are aged between 28-45
actively track their sleep (past 30 days)
suffer from subtle sleep fragmentation (e.g., >4 wake events/night)
show slight HRV declines which are not explained by illness or stress
From the result set of 430 users, it asks a selected target audience follow up questions:
“Have you experienced any recent memory lapses, attention issues, or word-finding difficulty?”
“Would you be willing to record a short voice message daily for the next 14 days?”
“Can we send you a 1-minute weekly focus/memory test?”
“Have you recently started or stopped any supplements or medications?”
The agent eventually develops concise instructions for individual users to help them with responding to these questions. The HPMP securely and asynchronously relays these prompts and their responses into the secure user data space and allows the requesting agent to query from them.
For selected users that share significant attributes, the agent can ask for further enrichment, e.g. to install applications that help collecting
Voice recordings (short prompts to analyze tone, cadence, word variety)
Typing patterns (to detect motor or processing slowdowns)
Weekly cognitive scores from in-app microtests
Advanced journaling data (mood, sleep quality, alertness ratings)
If the agent identifies a sufficiently correlated user cluster showing early cognitive biomarkers 5–7 years ahead of symptom onset, it updates a shared knowledge graph with multimodal neuro-risk signal nodes linked to demographics and observed habits and eventually anchors that finding as IP on chain (e.g. as IP-NFT). Here it also may include cryptographically provable provenance information that would allow for posterior contributor rewards or citations.
Depending on its instruction plan, the agent can add its new findings to a diagnostic knowledge graph that helps other application developers build screening tools for users with comparable bio markers. The agent can also directly return diagnostically certain findings, so the user in our previous example could be advised to seek medical consultation if they show signs that indicate a neurodegenerative issue. Ultimately, research agents help decreasing the time needed to incorporate new findings into third party algorithms that improve personalized therapeutics for users.
Users create profiles on the HPMP to collect their data from wearable devices and other health applications inside a space they ultimately control. The data is stored securely using user-owned encryption keys, and leverages technologies like Nillion's SecretVaults on jurisdictional compliant nodes.
HPMP User Profiles are publicly available to all implementers and will come with many benefits for third-party integrators:
They are enriched by MPC secured signers and transparently provide cryptographic key material and sufficient randomness to derive accounts for signing or authentication against blockchain networks, e.g., to control smart contract wallets for advanced use cases.
They unlock user controlled data encryption without any party beyond the MPC provider and the user having to keep any secrets safe.
Integrators gain many features of the protocol layer for free, and they stay related to the user base they helped onboarding into the protocol. Thus they can collect and read their users' information given their continued consent.
The protocol provides basic storage and accessibility information to integrators for free, including some technical guarantees over data storage safety, redundancy and compliance.
Research agents will never have direct access to the raw data lake. They instead interact with HPMP via MCP endpoints to find matching user profiles and establish secure communication channels with them. Agents can request access to specific data subsets from user profiles, such as readings from health wearables or aggregators like Apple HealthKit. Integrators that want to run research agents on their own user base will be able to do so at a major discount compared to external research agents.
Ultimately the HPMP user profiles allow agents to ask questions, deploy surveys on users' screens or present findings to actively gather feedback directly from individuals.
HPMP invites agentic AI developers, health aware users and 3rd party applications to use it as an openly accessible health data repository providing clearly defined access points to its profiles, user entities and their data. The HPMP never claims "ownership" of any user itself.
Our strategic vision for creating a self-sustaining cycle of user acquisition is as follows:
The DeSci ecosystem and personalized health application builders collectively recognize AI agents as the driving force behind research advancement.
Health application builders not only permit but actively facilitate users in exporting and synchronizing their information with HPMP infrastructure. In exchange, they are liberated from concerns regarding profile management or secure data storage at scale. Data collected on behalf of their users remains fully accessible for their operational requirements.
Agentic research findings flow back into knowledge graphs and discovery databases that are addressable by tool builders, so they can instantly update their applications and personalized health offerings.
Users benefit from tailored health insights fundamentally derived from findings to which they or their community members have directly contributed. This mechanism encourages more users to disclose their information to HPMP agents and engaging actively in conversations.
While not being a primary goal at time of the launch of HPMP, the protocol will require agents to pay for profile connectivity, queries and interactions with a token primitive that needs to be defined when the mechanism is laid out and has demonstrated a market fit with high confidence. The synergies between integrators, research agent ecosystems (like BIO) and external data providers are also yet to be determined.
Frankly, the HPMP will not put an immediate monetary incentive layer in place that simply would "pay users to share data" as this never turned out to be a reliable model; the immediate benefits users will gain for offering their information to HPMP matchmaking are:
They will receive very early hints on how changing lifestyle and behavior would contribute to certain conditions in ways that rarely have been tried before. Since their data becomes part of a shareable result set more quickly, the chances of getting early access to tools that help with individual medical conditions are far higher than waiting for startups or big pharma companies developing such tools on isolated siloes of their communities’ data.
The cryptographic nature of all HPMP entities will allow tracing the whereabouts and impact of individual information. Thus users can get access to early clinical drug trials that are promising candidates to help treat their conditions.
Users can decide to publish their own information to other protocols and researchers that refine algorithms on higher data dimensions - arbitrary wearable readings can help identify condition correlations that were unknown before; the HPMP's conversational subsystem can be utilized to reflect those findings back to data contributors.
A powerful new research paradigm is emerging and Welshare's HPMP will provide a crucial building block that keeps humans in the loop. Individuals decide which data they disclose and they receive real benefits in return: The gap between scientific research and affected individuals profiting from it has never been narrower.
The HPMP approach stands out for its commitment to user data sovereignty. Unlike plain data disclosure approaches where data is surrendered entirely, this system creates a secure environment where users actively participate in scientific discovery while their information remains protected by cryptographic technologies and only is decrypted inside trusted execution environments. This privacy-first approach encourages broader participation from diverse populations who might otherwise be hesitant to share sensitive health data, ultimately leading to more comprehensive and inclusive scientific discoveries.
The combination of AI-driven research agents, privacy-preserving storage protocols, end-user controlled data, long-term conversational memory and on-chain incentives will accelerate scientific discovery faster than anyone ever could have imagined.
Originally a studied mathematician (FH diploma), Stefan builds software of various dimensions and for many different purposes since 1999 - as founder, lead dev, specialist or simply as grizzly rubber bear for any team member who requires it. Since 2019 he focuses on products that value user sovereignty, privacy and permissionless backends. He has introduced many peers and community members to web3 stacks, p2p principles and decentralized thinking. In 2022 he joined Molecule as Senior web3 engineer and developed the specifications and smart contracts that enable Molecule’s IP-NFT and IPT ecosystem. After half a year of technical advising, Stefan joined welshare early 2025 to help starting up a protocol for truly sovereign and permissionless patient data utilization in an agentic research context.