Prism - AI trial matching, designed for clinical trust

The brief, and the setup around it.

Prism's design work began in late 2023, alongside the development of OncoLLM. The brief: turn Triomics' oncology AI into a product that cancer centers could run trial screening on - starting with the flows, interface, and design foundations.

Responsibility split: the founder owned discovery - user interviews, site relationships, demos at US cancer centers. Product managers owned PRDs. I owned the making: ideating flows, building POCs and prototypes, and shipping design iterations into the product.

Four constraints defined the work.

Domain complexity: oncology eligibility criteria are regulatory-grade material - staging, lines of therapy, biomarker thresholds. Designing screening interfaces required learning to read this material fluently.

Enterprise environment: Prism operates alongside EHRs (Epic and others) inside hospital workflows that cannot be redesigned. Clinical trust: LLM outputs are probabilistic, and users make decisions about real patients - every interface had to make evidence inspectable. Distance: users were at US hospitals, nine time zones away, with no direct access for research.

Requirements arrived second-hand. Each was traced back to its workflow.

Direct user research was not available in this setup. Inputs came through three channels: founder debriefs from site meetings and interviews, PRDs from product managers, and feature requests relayed from deployed sites. My method was to trace each requirement back to the workflow moment that produced it - who asked, during which task, and what happens before and after that moment - before designing against it.

Domain immersion covered the gap: trial protocols, eligibility criteria language, and ClinicalTrials.gov listings, studied until screening conversations could be followed without translation. Research was continuous across the two years rather than a discrete phase.

The product began as one flow: patient-to-trial matching.

Prism started as a single flow answering a single question: which trials fit this patient? One patient, their chart, a ranked list of trials, and the evidence behind each match. This flow was designed, prototyped, and shipped first.

The core interface decision was made here: evidence before verdict. An AI answer displays what it found and where, so the user can verify before acting. Later features inherited this structure - trial-to-patient matching is the same flow reversed; continuous tracking is the same flow re-run on chart updates.

Every idea was made clickable before it was discussed.

The working loop: requirement in → flow options ideated → clickable prototype built → founder presented it to sites → feedback returned → next iteration. Prototypes were the primary communication medium in every direction - with the founder, with PMs, with engineering, and (indirectly) with users.

Exploration happened through working flows rather than static concepts: multiple structures for a screening result were prototyped and compared directly. This kept decisions concrete and shortened alignment cycles in a team with no time for abstract design debate.

Feature by feature, under one mental model.

The platform grew request by request: trial-to-patient matching, cohort filtering, continuous tracking, notifications. Each arrived as an independent site request. The design task was integration - fitting each capability into one consistent screening model with shared states and repeating patterns, rather than shipping forty disconnected features.

Requests that broke the product's mental model were redesigned or pushed back on before implementation. This coherence discipline, maintained over two years of accretive growth, is why Prism operates as one product today.

Three operating decisions, with trade-offs.

Validation ran through production, continuously.

Validation was not run as moderated studies. Flows shipped into the live product, coordinators at deployed sites used them in real screening work, reactions returned through the founder and PMs, and subsequent iterations addressed them. This cycle ran for two years across hundreds of iterations.

The current platform is the accumulated result of that loop: every relayed issue and request absorbed into the system without breaking its core model.

Live, measured, and published.

Prism runs in production at leading US cancer centers - deployments publicly announced at Memorial Sloan Kettering, Yale Cancer Center, and Mount Sinai, among 10+ NCI-designated customers. Across sites, the platform screens 65,000+ patients for trials monthly. At a leading academic center: 7,000 patients screened monthly with 100% visit coverage, 3x the throughput of manual screening, a 30% lift in trial accruals, and 70% less time in eligibility review.

Published accuracy backs the workflow: 95.5% top-3 trial match accuracy (Nature Digital Medicine), 40% more matches generated (ASCO 2025), with criterion, reasoning, and citation accuracy each independently reported above 90%. The flows, prototypes, and shipped iterations behind these numbers are my work from the first patient-to-trial screen onward. Prism's deployment traction contributed to Triomics' Series B raise.

What deployed sites report.

From a leader at a partner cancer center, published on triomics.com: 'PRISM lets us cast a wider net than protocol eligibility alone, identifying patients earlier in their disease course, so we can offer trial options at critical points in their cancer journey. It has enabled us to pre-screen 100% of our patient visits against all our trials.'

The recurring theme in site feedback matches the original design intent: coordinators stop reading charts to rule patients out, and start their day with a cited, prioritized worklist instead.

A product built on an LLM, designed with AI in the loop.

Prism is a product that uses AI at its core - its answers come from OncoLLM, and LLM outputs are probabilistic: sometimes confident, sometimes uncertain, occasionally wrong. The interface work was largely about giving that behavior honest form - evidence attached to every answer, uncertainty represented as an explicit state rather than hidden, and re-runs, failures, and stale results all designed as first-class conditions instead of edge cases. Designing a product that uses AI is state design before it is screen design.

The workflow itself was AI-first. Claude handled requirement decomposition, flow ideation, edge-case enumeration, and interface copy; Figma and Figma Make turned flows into high-fidelity, testable prototypes in hours instead of days. This is what made 'hundreds of iterations' a real number rather than a figure of speech - the AI-assisted loop compressed each cycle from a sprint-sized effort to a day-sized one. Working on an AI product while designing with AI tools compounds: using the technology daily builds the intuition for designing it.

What I learned, and what it made me.

Process learnings first: process adapts to constraints. Without direct user access, founder debriefs and PRDs can function as research data if interrogated systematically. Without time for concept phases, prototypes can carry alignment. With a small design team and a platform-scale roadmap, coherence is maintained by holding and defending a single product model. The standing improvement item: closing the distance between designers and users - second-hand input built a working product; direct access would raise its ceiling.

Personal ones second. Prism gave me domain humility - I learned to read oncology protocols before presuming to design for the people who live in them. It built judgment under ambiguity: shipping decisions with incomplete information, then correcting fast without ego. It made systems thinking a habit rather than a phase - every screen weighed against the whole platform. And it settled what kind of designer I am: one who works where AI meets high-stakes workflows, fluent in the technology I design for and with. That's the practice I'm building a career on.