AI clinical research · Agent-to-agent

Hospital agents and pharma agents, running the trial together.

We don't design molecules. We prove them. Our agents work the same trial from both sides of the hospital wall. No patient record moves.

A2A interchange · protocol channel agents active

Hospital agents

Next to the source data.

protocol.parse chart.review cohort.screen eligibility.check ae.adjudicate visit.window deviation.alert sdv.auto

Sponsor agents

Next to the decision and the budget.

target.validate protocol.generate site.feasibility endpoint.select sample.size safety.signal causal.infer medical.write
Rule of the channel: records stay in the hospital, the model learns in place, an investigator signs before anything crosses.
5
of the global top-10 pharma companies pay
100%
renewal rate across those accounts
40+
agents in production at a leading tertiary GCP centre
<3 hrs
from protocol PDF to an EDC-ready database

Why this matters

Drug development still runs on three tens.

The clinical stage holds most of the time and most of the money. It is still run by people reading documents.

Fig. 01 · the arithmetic of one approved medicineclinical stage in teal

10 years to approval

Six are clinical. A year late costs a year of patent-protected sales.

6 of 10 years preclin I II III file
preclinical clinical review

$1B per approved drug

Over half is clinical. Site monitoring alone takes a quarter of trial cost.

$1B PER APPROVAL
clinical 55% preclinical 32% monitoring

1 in 10 reaches the market

Nine in ten never finish, usually only after running to the end.

PHASE I PHASE II PHASE III APPROVED 10 6.3 3.1 1
of every ten molecules entering the clinic

Everyone sells cost reduction. The real number is time, and it is two orders of magnitude larger.

Fig. 02 · what one Phase III decision is worthsame linear scale, $0 to $5B

Cost leverCut 20–30% off a Phase III

$30–60M

Time leverReach the market one year earlier

$1–5B

≈ 100× the cost lever · and a year of patients treated

A2A · agent to agent

Two agent networks, separated by a wall that should not come down.

Evidence sits in hospitals. Decisions and budgets sit in sponsors. We left the wall between them standing and put agents on both sides.

Fig. 03 · the A2A layerfive message types cross · four never do

HOSPITAL SPONSOR PROTOCOL BOUNDARY Chart review to structured data Continuous eligibility screening Deviation and visit-window checks Adverse event adjudication SOURCE SYSTEMS · READ IN PLACE HIS LIS PACS EMR CTMS Target validation and feasibility Protocol generation Site strategy and forecasting Causal analysis and filing WHERE THE DECISION SITS PIPELINE PORTFOLIO BUDGET FILING structured fields + provenance protocol constraints, queries eligible cohort counts feasibility: n=60 in 6 months? adjudicated AEs, endpoints Raw records · identified data · anything unsigned MODELS TRAIN IN PLACE · FEDERATED · CONFIDENTIAL COMPUTE never reaches the other side

Drag to see the full diagram

Crosses the channel

Structured fields, with provenance Eligible-candidate counts Adjudicated adverse events Endpoint statistics and safety signals Queries and amendments

Never crosses

Raw medical records Identified patient data Anything unsigned by the investigator Weights trained on unlicensed data

Automated source-data verification never writes back into the medical record. A person resolves every query.

The network

27 agents across the five stages of a trial.

Five are validated end-to-end in live, paid deployments. The rest are being validated one at a time.

01 · 5 AGENTS
Feasibility & selection
Target validation
Indication matching
Competitive intelligence
Strategic feasibility
Real-world evidence
02 · 6 AGENTS
Protocol design
Protocol generation
Eligibility optimisation
Endpoint selection
Sample size
Synthetic control arm
Virtual trial
03 · 6 AGENTS
Trial execution
Adverse event adjudication
Patient matching
System integration
Site feasibility
Recruitment forecasting
Real-time monitoring
04 · 6 AGENTS
Data & analysis
Statistical analysis
Causal inference
Biomarker discovery
Safety signal detection
Interim analysis
Draft study report
05 · 4 AGENTS
Submission & beyond
Submission dossier
Medical writing
Regulatory response
Submission tracking

Validated end-to-end in production. 22 specialty agents live at a leading tertiary GCP centre.

The platform

Let the protocol build its own database.

By hand, a study database costs a data manager three to four weeks. We parse the protocol once, and everything else falls out of it.

Fig. 04 · protocol to EDC-ready databaseone pass, no hand-written rules

T+0 T+3 HRS Protocol PDF AS WRITTEN 01 Structured parse Visits, I/E criteria, endpoints, safety rules 02 Field routing Each field bound to a source system and path 03 eCRF schema Tables, versioning and validation attached EDC-ready 517 FIELDS LIVE ZERO EXTRACTION RULES WRITTEN BY HAND · EVERY FIELD TRACEABLE TO ITS SOURCE NOTE

Drag to see the full diagram

Building it by hand

3–4 weeks

Read the protocol, design the tables, write the rules, debug.

On the platform

under 3 hours

One parse. Each square is a working day.

< 3 hrs
Protocol to EDC-ready
517
Fields routed automatically
0
Rules written by hand
100%
Field-level traceability

THGAgents

A causal reasoning engine for biomedicine

Traceable causal graphs, reasoning over paths, hypotheses that iterate. Built for what general models handle worst, literature that contradicts itself.

7.1Mpapers
12Mpatents
3,633targets mapped

AutoBioResearch · BioOpenClaw

A multi-agent network that keeps running

Generates hypotheses, verifies them in parallel, catches its own contradictions, and writes up once the evidence converges.

IJCAI 2026published
10+patents filed
4-layerarchitecture

Clinical rules live outside the model. The model only learns judgement. That is what makes the output explainable, portable, and safe to put in front of a regulator.

Evidence

Deployed, in production, and paid for.

Not pilots. Three constituencies, each with something at stake.

01 · Hospital

A leading tertiary GCP centre

40+ agents live against real trials, across six disease areas.

02 · Sponsors

Global top-10 pharmaceutical companies

Five of the world's ten largest buy from us, at 100% renewal.

03 · CRO

A supplier, not a competitor

CROs sell people. We supply agents, and we sell them to CROs too.

Fig. 05 · before / after

Recruitment window

Before12 mo
Now4–6 mo

Data-entry time

Before100%
Now−60%

90%+ precision and recall on structuring, patient matching up 41%, zero records out.

Fig. 06 · before / after

Protocol draft to sign-off

Before2 weeks
Now1 day

Investigator review effort

Before100%
Now−60%

First draft in 30 minutes, 99% accuracy on three-way verified review, 100k+ scattered preclinical records made queryable.

Fig. 07 · before / after

Adverse event report processing

Before60 min
Now5 min

0

people supplied

Priced per field of output, not per person-month or per visit.

The moat

Five capabilities. Take one away and none of it works.

Not a list of strengths, a list of gates. Each one has stopped a well-funded competitor.

Five interlocking capabilities required to run agent-to-agent clinical trials All five, or none. 01 HOSPITAL ACCESS 02 DATA GOVERNANCE 03 DRUG · MED · AI 04 PI SIGN-OFF 05 SPONSORS PAYING
  1. 01

    Getting into the hospital

    Six to twelve months through IT, ethics and clinical review. Live at a leading tertiary GCP centre.

  2. 02

    Compliance and data governance

    Data cannot leave, and structuring free text is unglamorous work. Models train locally at 90%+ precision and recall.

  3. 03

    Drug development, medicine and AI at once

    Each field locks the other two out. A pharma R&D veteran, a Stanford AI and ex-Google scientist, and senior advisors, in one team.

  4. 04

    The investigator's trust

    If the investigator does not sign, the AI has done nothing. Review effort down 60%, 99% verified accuracy.

  5. 05

    Sponsors who actually pay

    Twelve to twenty-four months of procurement, legal and audit. Five of the global top ten pay, at 100% renewal.

Only AI, and you never get through the door. Only hospital relationships, and you never ship. No investigator, no sponsor, and it was only ever a demo.

Who we work with

Sponsors on one side. Hospitals and universities on the other.

This only works if both ends are real.

Pharmaceutical partners

MSD
Global top-10 sponsor
Pfizer
Global top-10 sponsor
AstraZeneca
Global top-10 sponsor

Academic & hospital partners

Peking University
Clinical research
Duke University
Clinical research
Huazhong University of Science and Technology
Clinical research

The hospital brings

Clinical setting Research demand Data environment System interfaces Principal investigators The ethics pathway

We bring

The platform Clinical research data lake Governance engine Adjudication models Implementation and operations

The hospital gets research capacity and GCP throughput. We get the right to develop against the data. Both are written in before anything is installed.

Where this goes

From replacing the manual work, to running the trial before it starts.

Every trial leaves data behind, real records plus the judgement of the physician who corrected the draft. Each step moves us earlier.

Fig. 08 · where each capability actsthe intervention moves earlier

BEFORE THE TRIAL DESIGN ENROL RUN READ OUT LIVE TODAY AI trial platform Agents replace the manual operating work on both sides IN BUILD Digital twin patients What if this patient had taken the other drug? NEXT Synthetic control arms 30–50% fewer patients, 6–12 months less THE POINT OF ALL OF IT Virtual trials Run it a thousand times in silico, so the failures fail there VIRTUAL TRIAL ENGINE CHECKED AGAINST A COMPLETED PHASE III · WITHIN ABOUT TWO PERCENTAGE POINTS OF THE REAL RESULT

Drag to see the full diagram

Still openThe omics cohort has to be fully connected.
Still openRegulatory acceptance of synthetic controls is narrow. Pre-submission discussion, not approval.
Still openWorking at one hospital is not reproducing it at thirty.

Team

Drug development, medicine and AI, in the same room.

A hard team to assemble, which is most of why it works.

Founder & CEO

Xincheng Zhang

  • MS in Artificial Intelligence, Stanford University
  • Previously search and recommendations at Google; co-founder and CTO of QTC Care
  • SC09 international supercomputing champion · Forbes 30 Under 30, healthcare

Co-founder & Chief Scientist

Kevin Jin

  • A decade in innovative-drug R&D and clinical AI
  • Among the first in China to put large language models into production in biomedicine
  • Worked on the pipeline behind the world's first AI-designed drug to reach Phase III

Medical advisor

Twenty years in multinational pharma. Previously head of medical affairs at a top-10 company.

Scientific advisor

Chair professor of cell biology at a leading European university. An authority on Alzheimer's and ageing.

Hospital partnerships

Fifteen years of hospital partnerships. Previously an executive at a major hospital IT company.

Our vision

Better medicines, reaching the patients who need them, sooner.

Every month a trial runs longer is a month someone waits. Everything else on this page is a way of getting to that number.