Evidence infrastructure for clinical trialsPsychiatry first, every area next

Faster, cheaperclinical trials.

We rebuild half of a trial’s control group from real medical records. Sponsors recruit fewer patients, spend less, and reach a readout sooner. Benchmarked first in psychiatry, where the outcomes live in clinical notes.

Book a call See the evidence

One control arm of a late-stage trial, 1,000 patients, 25 to a block. Footprint counts patients, height is what each one costs.

Conventional

With Sapien

1,000 recruited

500 recruited

500 rebuilt from records

The model reads EHR, Clinical notes, Labs, Imaging

500 patients you never recruit$40k+ saved on each

Same control arm. Half the recruitment.

Two grids on one ground, each the same control arm of 1,000 patients. Footprint counts patients and height is the cost of one patient. The conventional arm is a full grid of recruited blocks, ten columns of one hundred, four rows deep, twenty-five patients to a block. The Sapien arm is five of those columns recruited at the same height, then a lilac handover plane standing at the five hundred line and split into one fin per record source, then five columns rebuilt from records, tessellated into facets and lying a tenth as tall because a rebuilt patient costs a tenth as much. The open periwinkle cage over the rebuilt half is the recruitment that never happens.

The problem

The control group is the bottleneck.

A late-stage trial recruits about 2,000 patients, and half receive a placebo. Enrolling those control patients is the slowest, most expensive part of the study.

$50k+Cost per control patient

To recruit one patient the traditional way.

37%Sites that miss enrollment

Fall short of their enrollment target.

11%Sites that enroll no one

Open, then never recruit a single patient.

We have bought datasets for over $70M for other studies. For this, they are practically useless.

Director of Biostatistics, global pharma company

The solution

Recruit half. Rebuild the rest.

We rebuild half of the control group from real medical records, as trial-grade synthetic patients. The sponsor recruits fewer people, and every later trial reuses the cohort at lower cost.

Data supply

IndiaLive

Two hospital groups

EuropeMoU signed

Twelve-hospital chain

United StatesIn progress

Late-stage conversations

Enrichment

One super patient

Scattered encounters resolved into one longitudinal patient timeline.

Merged fromEHRClinical notesLabsImaging

Output

Reusable cohort

Assembled once, drawn on by every trial. Cost falls with each one.

Matched to your inclusion and exclusion criteria, rather than rebuilt for each study.

Data stays on-site. Records never leave the partner’s servers.

Trial revenue is shared back with partner hospitals.

Control patients to recruit1,000from 1,000

Est. control cost$50.0M

Recruitment18 mo

Full control arm · all recruitedRecruited

External control · built from records

One glyph = 6 patients · Illustrative

$40k+Saved per patient

Every control patient we rebuild instead of recruit.

50%Of the control group replaced

Rebuilt from real records, item-level traceable.

10xLower cost

For the portion of the arm we reconstruct.

Evidence

Measured against frontier models.

Our reconstruction method is described in a public paper and measured on the Psych-ECA benchmark. On semi-synthetic data its prediction bands covered the true outcome 93 to 96 percent of the time, at or above the nominal 90 percent target.

See the evidence Read the paper

Psych-ECAarXiv 2607.27224

93–96%

Empirical coverage of nominal 90% prediction bands. Gradient boosting reached 87 to 88%. Benchmark results on semi-synthetic data, not clinical validation.

Product

See the control arm, patient by patient.

Navigator makes a rebuilt external control arm inspectable before it ever enters a study, down to the individual patient. This is a browser recreation that uses fictional but plausible data and reduced interface detail.

Follow one rebuilt patient's recovery week by week, and decide whether they are credible enough to stand in for a recruited control.

ELEVATE-TRD Phase 2bgenerated

v1.3 · n=847

Patient SYN-WBTQR5

65+FLatin AmericaResponder

MADRS depression severityGAF global functioning

26Baseline MADRS

6.3Week 8 MADRS

-76%Change

847PatientsSynthetic control arm

46.3%Week 8 Response392 of 847

9.6%Week 8 Remission81 of 847

Baseline MADRS: 32.9 ± 4.4
Baseline GAF: 49.6
Illness Stage: 33% chronic

Patient ID	Age	Sex	Illness Stage	BL MADRS	Wk8 MADRS	Change	Responder	Prior Lines
SYN-WBTQR5	67	F	Recurrent	26	6.3	-76%	Yes	2
SYN-K3PLXA	41	M	Chronic	34	29	-15%	No	4
SYN-M9DRTC	29	F	First-episode	31	12	-61%	Yes	1
SYN-Q7VHNE	55	M	Recurrent	38	33	-13%	No	3
SYN-B2XKUF	34	F	Chronic	36	31	-14%	No	5
SYN-Z5TWMD	62	F	First-episode	28	13	-54%	Yes	1
SYN-R8NCJP	47	M	Recurrent	33	26	-21%	No	2
SYN-H4GBVS	38	F	Chronic	40	34	-15%	No	6
SYN-L6YQAN	51	M	First-episode	30	14	-53%	Yes	2

The marked patient is the one featured in the Trajectories view. Synthetic records, illustrative values.

Copilot

Ask the cohort a question in plain language and get the answer in the same view where you decide, not in a separate export.

Why now

The precedent is already there.

Regulators have accepted external and in-silico controls for a decade. The methods only recently became good enough for psychiatry, where the outcomes are subjective.

201621st Century Cures Act, RWE program
2022FDA Modernization Act 2.0, AI and in-silico methods
2023FDA draft guidance, externally controlled trials
NowThe technology is ready for psychiatry

The pressure

About $170B of drug sales lose patent protection by the early 2030s. Sponsors need cheaper, faster trials to keep pipelines full.

The opening

Psychiatry came last because its outcomes live in notes. That is exactly the problem our method was built to solve.

Team

ShashankCEO

IIT Delhi. Biotech, gene therapy, and microfluidics.

Aakash BhagatCTO

BITS Pilani. Machine learning for Turing-award labs in the US.

Advisors affiliated with Pfizer, GSK, Tufts, Johns Hopkins, Imperial, and UCL.

Fewer patients.
Faster readout.

A short call to map your program onto a rebuilt control arm. No raw records leave the source.