Skip to content
System build

A clinician evidence platform, surveillance to point of care: federated search over scattered biomedical sources, one deduplicated stream.

Client
LitCurrent
Industry
Biomedical research
Location
Washington DC
Problem

Clinical evidence lives across PubMed, Europe PMC, OpenAlex, and ClinicalTrials.gov — four databases with different interfaces, overlapping records, and paywalled full texts. Researchers lose hours to the same paper under different identifiers, and a review built on an incomplete corpus is a review with a hole in it.

What was built

The system.

  1. 01

    LitPortal

    Point-of-care evidence search with citation-verified synthesis: the federated multi-agent engine queries the sources in parallel, reconciles duplicate records, and returns a single ranked stream.

  2. 02

    LitCurrent

    The evidence-awareness workspace: literature surveillance, an article library, audio learning, and AI-supported analysis.

  3. 03

    Platform

    Four backend services behind one login and one shared library, alongside the wider modules: LitScreen, LitHub, LitStream and LitScholar.

Engineering decisions

What was chosen, and what it cost.

  1. 01

    Four services instead of one monolith: two application backends (LitCurrent and LitPortal) beside shared Identity and library platform services.

    Tradeoff

    Four repositories and an internal service-auth layer to operate, instead of one deployable.

    Result

    One login and one canonical article library across both applications, with each backend still independently deployable.

  2. 02

    Stateless auth: the identity service issues one signed RS256 token and every service verifies it offline against cached JWKS keys.

    Tradeoff

    Revocation propagates on token expiry rather than instantly, and key rotation has to be managed with care.

    Result

    Authentication adds no cross-service round trip to any request.

  3. 03

    Saves mirror into the shared library through a durable outbox with scheduled retries, not an inline call.

    Tradeoff

    The library is eventually consistent — a save can land seconds later — and a dead-letter path has to be operated.

    Result

    A save survives a downstream outage, and every mirror attempt is auditable.

  4. 04

    AI interprets, deterministic systems decide: retrieval, ranking, source eligibility, article selection, and citations are controlled by deterministic code. The model may interpret clinical context and surface labelled supplemental candidates — nothing more.

    Tradeoff

    Gives up the convenience of end-to-end LLM answers; every AI-adjacent path needs its own enforcement gate.

    Result

    Citation-verified answers a clinician can check — the platform cannot hallucinate a PMID.

  5. 05

    A hybrid data layer: document stores for the application backends, PostgreSQL for identity and the canonical library.

    Tradeoff

    Two database technologies to operate, and a consolidation path that must be planned rather than assumed.

    Result

    The one-entry-per-user-per-paper guarantee lives in a relational core while application workflows keep their flexible stores.

In the product
LitScreen: the scheduled literature feed with abstracts, AI summaries, and save-to-library screening
LitScreen: scheduled surveillance — the digest feed screened article by article, AI summary beside the abstract, one tap to the library.
LitHub: the saved-article library with abstracts, AI summaries, audio, LitScholar and LitClub actions
LitHub: the canonical library — every saved article carries its abstract, AI summary, audio, deep-dive and discussion actions.
LitStream: generated audio summaries of saved articles with play and download controls
LitStream: audio learning — saved articles become playable, downloadable summaries with specialty tags.
LitPortal: a clinical question answered with deduplicated, ranked evidence in a sortable table
LitPortal: point-of-care search — a clinical question resolved through the PECO framework into 22 deduplicated records, ranked and filterable.
Delivery

Built as a fixed-scope engagement, live in production. A separate five-week scope merged LitPortal into the platform, moved hosting under the owner's control, and produced the handoff package for the downstream migration partner.

Outcome

What changed.

Deduplication accuracy
99%+Deduplication accuracy

across records federated from multiple biomedical sources, where the same paper appears under different identifiers.

Backend services, one login, one shared library
4Backend services, one login, one shared library

two application backends beside shared identity and library platform services.

Initial search response
<2sInitial search response

for a federated query spanning multiple upstream sources.

Deployed in production
LIVEDeployed in production

running for clinicians, not a prototype.

Full-text recovery
>40%Full-text recovery

improvement in retrieving full text for records where it was previously unreachable.

Related workAll work →