observed signalstable identifiers and exact field agreement
product decisionresolve with deterministic rules
how it was checkedfield-level checks and acceptance testing
pharmaceutical data-matching platform
A hybrid matching workflow that reconciled more than 50,000 pharmaceutical records while keeping uncertain cases visible and auditable.
50k+
95%+
<2 min
≈90%
≈24h
overview
The product standardized source data, resolved common cases through deterministic rules, escalated ambiguous records to an LLM fallback, and preserved human review where confidence was insufficient.
the user — Research managers reconciling pharmaceutical records across inconsistent sources.
the problem
The existing process required manual cleanup, comparison, and validation across tens of thousands of records. Common matches consumed the same attention as genuinely ambiguous cases, while incorrect matches weakened downstream research.
what i owned
- mapped the reconciliation workflow and defined measurable success criteria
- translated research and legal constraints into product requirements
- shaped the deterministic-first matching strategy with engineering
- prioritized validation, exception handling, and export workflows
- coordinated acceptance testing across product, research, and legal stakeholders
evidence and decision sequence
observed signalpartial agreement or inconsistent naming
product decisionroute to the LLM fallback
how it was checkedcompare proposed match with source evidence
observed signalconflicting identifiers or unresolved ambiguity
product decisionhold for human review
how it was checkedresearcher accepts, corrects, or rejects
evaluation framework
optimize for trustworthy automation, not maximum automation
failure modecostproduct responsefalse positive match
incorrect data enters downstream research
require stronger agreement or review
false negative matchvalid match remains manual
route ambiguity without auto-accepting it
unresolved conflictautomation cannot justify a decision
preserve source evidence for human review
key decisions and trade-offs
use deterministic matching before an LLM fallback
alternativesUse an LLM for every record, rely only on fuzzy matching, or keep ambiguous records manual.
what we choseDeterministic rules handled common cases; an LLM was reserved for ambiguity.
whyThe hybrid approach improved speed, predictability, auditability, and operating cost.
trade-offMaintaining explicit rules added complexity but reduced LLM token usage by approximately 90%.
surface uncertainty instead of hiding it
alternativesAutomatically accept every output or require review for every match.
what we choseUse field-level statuses and route only low-confidence cases to human review.
whyResearchers retained control without recreating the full manual workflow.
trade-offThe interface needed more states, but trust and exception handling improved.
product interface
matching workbenchsanitized portfolio reconstruction
source record
drug namecardiovex
exactingredientmetoprolol
exactstrength50 mg
exactdosage formtablet
partialapplicationANDA 20481
exactdeterministic matchconfidence 0.98matched automatically
candidate record
product namecardiovex
exactactive ingredientmetoprolol
exactstrength50 mg
exactdosage formoral tablet
partialapplicationANDA 20481
exactambiguous example dosage form and application conflict
rules inconclusiveLLM fallbackhuman reviewresults
workflow
Approximately 24 hours of recurring manual review were removed each month.
automation
More than 95% of queries were resolved automatically in under two minutes.
efficiency
The deterministic-first strategy reduced LLM token usage by approximately 90%.
measurement notes
Definitions shown so the claims can be discussed and challenged.
95%+ automated resolution
Share of reconciliation queries resolved without recurring manual review during acceptance testing.
≈24 hours saved monthly
Estimated from recurring review volume multiplied by the prior average manual reconciliation time.
≈90% lower LLM usage
Compared with routing every record through the LLM; deterministic coverage removed common cases first.
under two minutes
Observed end-to-end runtime for the 50K+ record workflow in the shipped implementation.
reflection
The strongest product decision was not adding more AI. It was defining precisely where rules were sufficient, where AI created value, and where human judgment still mattered.