How TrialMatch built voice-first clinical trial matching using Vocal Bridge

From a hackathon idea to a company solving one of healthcare's most consequential matching problems, built on Vocal Bridge's Voice for App surface.

The problem

A cancer diagnosis triggers an avalanche. Within days, a patient, newly disoriented and sometimes still in shock, is navigating oncologist appointments, treatment decisions, insurance questions, and a wall of medical terminology they have never had to learn before. Somewhere in that chaos sits the question of clinical trials, a potential lifeline that, for most patients, never gets seriously evaluated.

The numbers are stark.¹ Only 5% of cancer patients ever join a clinical trial, though 75% would enroll if someone told them about it. 80% of trials miss enrollment targets, costing the industry an estimated $8 million per day in delays and roughly $1,200 per failed patient screen.

The matching infrastructure fails patients three ways. Oncologists have 15-minute appointments and no incentive to refer. The 400,000-trial corpus on ClinicalTrials.gov uses dense medical jargon with 30+ eligibility criteria per trial. And phone-based coordinator screening does not scale. TrialMatch's bet is that voice is the only intake modality that scales, because a conversation, with an agent that elicits and clarifies in plain language, is the only interface a newly diagnosed patient can actually navigate.

They bet on voice, and they bet on Vocal Bridge to make it happen.

What TrialMatch built

TrialMatch is a voice agent that runs clinical-trial matching as a guided conversation. A patient talks through their medical history (cancer type, stage, prior treatments) and constraints like travel distance, and the agent searches active trials, evaluates eligibility, surfaces specific options at named cancer centers, and offers to go deeper on any one of them. A full intake takes a few minutes, by voice, with no forms to fill out.

The goal is to make trial matching something a patient can actually do right after a diagnosis, when a form is the last thing they can face. Voice fits because eligibility is a conversation rather than a filter. The agent can ask the follow-up a form never would, adjust the search when a criterion rules a trial out, and explain why a given trial does or does not fit.

The accuracy that matters

TrialMatch reports its matching engine validates against 87%+ accuracy on NIH benchmarks, well above where search-website matching and human coordinator screening land. The system evaluates every eligibility criterion for each candidate trial rather than filtering on a small set of fields. That is what makes the "we found these but they don't fit your treatment history" moment in the demo work. The agent is not filtering. It is evaluating.

How TrialMatch uses Vocal Bridge

TrialMatch is built on Vocal Bridge's Voice for App surface, the integration pattern where the voice agent and the UI talk to each other on the same WebRTC data channel. When the agent extracts a symptom, the eligibility profile on screen updates. When the patient corrects an entry on screen, the agent sees the change and incorporates it into the next search.

The architectural split. TrialMatch owns the matching agent, the eligibility evaluator, the 400,000-trial corpus, and document processing for pathology and labs. Vocal Bridge owns the voice channel, the real-time WebRTC transport, streaming-aware text-to-speech, barge-in, transcript streaming, and the conversational fillers ("give me just a moment") that keep a patient grounded while a tool call runs.

import { useAgentActions } from "@vocalbridgeai/react";

const { onAction, sendAction } = useAgentActions();

// agent -> app: an extracted entity updates the visible eligibility profile
onAction("add_condition", ({ condition, icd10 }) => {
  setProfile((p) => ({ ...p, conditions: [...p.conditions, { condition, icd10 }] }));
});

// app -> agent: the user corrected something, tell the agent
const handleRemove = (id) => sendAction("remove_condition", { id });

Illustrative. The actual TrialMatch implementation handles richer medical entities.

"Built a Codex plugin for Vocal Bridge and submitted my YC startup entirely on the platform." Joel Wang, founder of TrialMatch

Joel contributed his Codex plugin back to the community stack, an extension other Vocal Bridge builders can now use.

What's hard about voice here

A voice intake only works if it can hold a long, branching medical conversation without feeling slow or rigid, so a build like this has to clear a few real bars.

• The agent runs searches mid-conversation. Querying live trials, and re-querying when a criterion rules one out, takes a beat, and dead air makes a frightened patient think it broke. Vocal Bridge's streaming, low-latency speech keeps the reply moving while the search runs.
• Patients correct themselves out loud. Someone will cut in to fix a detail ("it was immunotherapy, not chemo") and expect the agent to absorb it and adjust. Vocal Bridge's barge-in handles that turn-taking, so the team did not have to build it.