Decision support for life-sciences MLR

The review that used to take a week now takes minutes

Pharmaceutical MLR teams still reconcile every promotional PDF against an Excel claims matrix, the required ISI, and a reference list, by hand. ARK runs those three checks in one pipeline and attaches evidence to every flag.

Sign in

≤ 2 min per document · 100% holdout precision · full decision trace · managed secrets

review · asset_q3_hcp.pdf87s

Page 2 of 6

Findings · 3 of 14

Ad-hoc claim

“Clinically proven fat reduction” is not in the approved matrix.

Uncited claim

Efficacy statement on p.2 has no resolving citation marker.

ISI verified

Required safety block matches the consumer / digital template.

The problem

Three checks. One spreadsheet. A calendar week.

A reviewer opens the PDF, the approved claims matrix, the ISI required for that product and tactic, and the reference list. They hunt for off-matrix claims, missing safety language, and citations that do not resolve. One miss delays a launch or invites a letter. At 20 to 40 assets a week, the work does not scale.

Promotional PDF

  • Headline claim
  • Body copy · p.2
  • ISI block
  • References

Claims matrix.xlsx

  • 412 approved rows
  • Audience · qualifier
  • Off-matrix = reject

ISI + citations

  • HCP vs consumer
  • Channel × tactic
  • Does ¹ resolve?

Human MLR first pass

Ctrl+F, email threads, 5 to 7 days in flight

  • Claims vs. the matrix

    Every promotional sentence is hunted against hundreds of approved rows in Excel, including wording, audience, and qualifiers.

  • ISI vs. the required template

    Safety language must match the template for product, audience, channel, and tactic. Multi-product assets multiply the checklist.

  • Citations that actually resolve

    A superscript that does not land on a real reference is still a finding. Reviewers walk the list by hand.

How we solved it

Parse once. Retrieve the matrix. Reason once. Leave a trace.

ARK 2.0 is a measured pipeline, not a chatbot pasted onto a PDF. Datalab parses the asset once. Hybrid retrieval finds the closest approved claims before the model speaks. The model reasons once. Every finding keeps a paper trail a reviewer can defend.

Stage 01 · Ingest

PDF lands in object storage. FastAPI validates, writes the document row, and enqueues an ARQ job on Redis.

Hybrid matrix retrieval. The model never starts cold.

Exact
Fuzzy
Voyage vector
Rerank top-5

The value

Hours back to the reviewer. A paper trail for the record.

ARK is decision support. A human still signs off. What changes is the first pass: measured, evidence-linked, and fast enough to keep up with 20 to 40 assets a week.

Holdout set · May 2026 · Gate 2 passing

0%

Precision

0.0%

Recall

0.0%

F1

Scored on a held-out validation split that is never used to tune prompts. Production target remains ≥95% precision and recall. The harness existed before the pipeline did.

  • Manual first pass

    Before

    5 to 7 days

    After

    minutes

    Reviewers used to reconcile PDF, Excel, ISI, and refs in flight.

  • Vs. ARK 1.0 processing

    Before

    5 to 10 min

    After

    ≤ 2 min

    Design targets: p50 ≤ 90s, p95 ≤ 180s end-to-end.

  • Reasoning architecture

    Before

    6 agents

    After

    1 pipeline

    One coordinated pass plus targeted validation. No fragile handoffs.

Why it is different

Not a chatbot on a PDF. A retrieval-backed review engine.

Generic models summarize. Spreadsheets do not scale. ARK 1.0 fragmented the job across six agents. This stack retrieves the approved matrix first, reasons once, and stores the evidence.

Claims vs approved matrix

ManualCtrl+F through Excel
Generic LLMGuesses from the PDF
ARK 2.0Hybrid retrieval + matrix match

ISI vs required template

ManualChecklist by tactic
Generic LLMNo template binding
ARK 2.0Product × audience × channel × tactic

Citation integrity

ManualWalk the superscripts
Generic LLMMisses layout structure
ARK 2.0Marker must resolve in-document

Evidence / audit

ManualEmail and memory
Generic LLMA paragraph of prose
ARK 2.0Page, bbox, decision_trace

Processing time

ManualDays in flight
Generic LLMFast, untrusted
ARK 2.0≤ 2 min, measured p50 / p95

Architecture

ManualPeople and spreadsheets
Generic LLMOne prompt, no harness
ARK 2.0Parse once · reason once · score always

The stack. Built to wait efficiently on parse, embed, and reason.

Interface

  • Next.js 16
  • React 19
  • Tailwind + shadcn
  • pdf.js workspace

API & workers

  • FastAPI · Python 3.12
  • ARQ on Redis
  • Pydantic v2
  • async end-to-end

Intelligence

  • LLM reasoning
  • Voyage embed + rerank
  • Datalab Accurate
  • Langfuse traces

Data & ops

  • Postgres 18 + pgvector
  • S3-compatible storage
  • Railway
  • Managed secrets

System context. Reviewers stay in ARK; models stay behind the API.

Reviewers · agencies · admins

ARK 2.0 platform

Next.js workspace → FastAPI → ARQ workers → Postgres + object storage

  • Datalab Accurate
  • LLM reasoning
  • Voyage AI
  • Langfuse

The three MLR checks, in one workspace

Built for teams who live in the claims matrix, ISI footnotes, and reference lists, and need every flag to jump back to the source.

  • Single AI pipeline

    One coordinated review pass plus targeted validation. Fast, consistent results without fragile handoffs between tools.

  • Claims matrix alignment

    Every promotional claim matched to approved copy or flagged as ad-hoc, with confidence and source text preserved.

  • ISI you can defend

    Extract ISI blocks and compare to the required template for product, audience, channel, and tactic, including multi-product assets.

  • Reference integrity

    Verify citations for claims that need substantiation and surface references that never appear in the body.

  • Traceable findings

    Click any flag to jump to page and paragraph. Accept, reject, or comment, then export a clean report for the record.

  • Modern parsing

    Structured document ingestion with durable storage of originals and parsed output for replay and audits.

From upload to a defensible sign-off

Agency submits, the pipeline runs, the reviewer decides. Designed for p50 ≤90s and p95 ≤180s processing. The human still owns the approval.

01

Ingest

Upload PDFs up to 100 MB, text-native or scan-ready. Originals and parsed structure are stored for audit.

02

Analyze

High-accuracy parsing and coordinated analysis extract claims, ISI, and references in one pass.

03

Review

Side-by-side document and findings. Jump to evidence, resolve flags, and collaborate with clear ownership.

04

Report

Export asset-level reports and feed dashboards for trends by product, channel, and rejection reason.

Security that passes IT review

ARK 2.0 is designed for environments where marketing operations meet strict infosec and compliance expectations.

Enterprise auth

SSO via SAML or OIDC with role-based access: reviewer, agency, admin, and read-only. Every action is attributed in the audit log.

Secrets & data

No hardcoded keys: credentials live in a managed secret store. Encryption in transit (TLS 1.3) and at rest (AES-256).

Logging and traces

Structured logs, request tracing, and a named audit log for reviewer and admin actions. Every finding keeps a decision_trace.

Bring the matrix, the ISI, and the citations into one pass

Same regulatory rigor. Minutes instead of a week. A decision trace reviewers can stand behind. ARK does not approve assets on its own.

Sign in