Pharma Data Pipeline
- Preserves source artifacts and native identity
- Retains document and version context
- Builds source-bound read models
- Exposes structured, inspectable access
Selected work · eMonograph + Pharma Data Pipeline
I turn complex pharmaceutical evidence into products people can trust and use.
An end-to-end product system spanning regulatory source ingestion, evidence modeling, human-centered exploration, and a governed AI workflow strategy.
Jay Sethi·Product strategy, architecture & implementation·jsethi.com
I didn’t build another pharma chatbot. I built the evidence layer it should stand on — so every answer can retain its product, market, document version, source section, and exact supporting passage.
The product thesis: make evidence addressable before making AI persuasive.
Independent product lead + builder
I worked across the entire product system.
- Framed the opportunity around evidence preparation — a valuable, testable job before autonomous content generation.
- Designed the domain model for product identity, source versions, evidence bindings, interpretation, and review state.
- Architected the experience that turns retained regulatory documents into explorable, source-resolved information.
- Directed AI-assisted implementation across pipeline, Reader, read-only API, MCP surface, and qualification harnesses.
- Built the proof discipline that separates processing success, semantic support, workflow permission, and business impact.
Strategy, experience architecture, systems thinking, and hands-on delivery — one coherent practice.
Built foundation · designed activation path
From raw regulatory sources to decision-ready work.
eMonograph Reader
- Searches products in source context
- Navigates structured label sections
- Inspects exact supporting passages
- Keeps review work attached to evidence
Activation blueprint
- Defines how to scope the communication job
- Designs review-ready evidence packages
- Makes unsupported requests visible
- Maps the path to controlled downstream reuse
The implemented pipeline and Reader establish one evidence contract; the activation blueprint extends that contract into a measurable commercial workflow.
Product decisions · select to expand
The architecture turns rigor into product leverage.
Drug name, regulatory product, package, document, and version remain distinct. The system can resolve the right thing before it summarizes anything about it.
Original artifacts and native structure stay separate from extracted values and semantic mappings. Parser repairs can improve the product without rewriting the source.
Not assessed, unavailable, not observed, conflicting, and explicitly absent are different states. That gives users honest next actions instead of a misleading empty field.
Source integrity, semantic support, reviewer disposition, and distribution authorization are separate gates. The same foundation can support exploration today and governed reuse later.
Engineering judgment · select to expand
I found the failure modes that polished demos miss — and fixed them at the contract level.
A BRAFTOVI fixture exposed an ingredient field populated by a chemical-name fragment. I strengthened the admission rules, bound the non-proprietary name to the correct passage, separated route, dosage form, strength, and package count, and prevented indication-specific status from becoming product-wide status.
A citation existed but could re-resolve through a duplicate-prone section label. I preserved the originating section-row identity and constrained resolution to the correct version, turning “has a source” into “points to the passage that actually supports this value.”
I instrumented a 10,000-record replay, traced a roughly 3.1 GB peak to eager all-pages PDF extraction, then moved to page streaming and explicit cache release. Post-fix peak: 383.7 MB, with zero processing failures and the database unchanged.
The pattern: trace the defect to the system boundary, strengthen the contract, and add the check to qualification.
Recorded full-corpus qualification
Built and tested at real regulatory-data scale.
292,649
records completed
30,718,387
source-evidence bindings checked
0
processing failures
Source families271,038 FDA SPL
21,611 Health Canada DPD
21,611 Health Canada DPD
Frozen databaseSHA-256 unchanged
Replay loop27.9 hours · 2.92 records/sec
Semantic fixtures5 / 5 curated cases passed
Evidence checksIdentity, artifact, version, quote digest
Operating boundaryRead-only dry replay
Report28 September 2026
Data effectNo canonical data materialized
This establishes source-processing and binding integrity at scale. Broader semantic review remained a separate release gate — exactly the distinction the product is designed to preserve.
Experience architecture
A product surface, not a data dump.
1shared evidence contract across the pipeline, Reader, API, and agent tools.
- Find the right product in its regulatory and jurisdictional context.
- Navigate structured sections instead of hunting through long source documents.
- Inspect the exact passage behind a displayed value.
- Compare supported contexts without silently merging identity.
- Carry evidence forward into briefs, review questions, and controlled workflows.
The interface makes provenance useful to people; the API makes the same contract useful to software.
Where the foundation creates value
Five high-friction jobs become one connected workflow.
Evidence discovery
Label-backed briefs
Claim-reference packages
Reviewer questions
Controlled reuse
The entry point is deliberately narrow: one brand, one jurisdiction, one HCP communication job. That makes value measurable and adoption practical.
“A shorter path to a reviewable answer, with the evidence still attached.”
Interactive workflow
Turn a scoped marketing question into a review-ready evidence package.
Product strategy: land with evidence preparation, prove the job-level value, then expand into approved reuse and platform handoffs.
A clear place in the stack
Differentiated where the work is hardest: before content enters approval.
What established platforms do well
Systems such as Veeva PromoMats manage claims libraries, linking, modular content, and formal review workflows.
That is the downstream system of record — not something eMonograph needs to imitate.
What eMonograph adds
An upstream evidence-preparation layer that resolves source identity, preserves version-specific passages, exposes uncertainty, and hands reviewers a stronger package.
The commercial wedge is better prepared inputs, fewer context-rebuilding steps, and cleaner integration into the sponsor’s existing approval process.
My product approach
Land a bounded result in six weeks — then earn the right to expand.
Define the proof
Choose one brand and 20 representative historical tasks. Lock source rights, reviewer ownership, manual baselines, and the evaluation set.
Run real work
Execute comparable manual and assisted tasks. Capture active preparation time, package completeness, abstentions, and correction effort.
Make the decision
Review independently where practical, replay seeded changes, reconcile results, and decide whether to expand, revise, or stop.
Decision-ready, not demo-ready
Set thresholds with the sponsor before testing: preparation time, critical identity or unsupported-claim defects, qualifier retention, and reviewer correction burden.
Illustrative go/no-go target from the proposal: 30% lower median active evidence-package preparation time, with no observed critical defects in the scoped sample and no increase in reviewer correction burden.
What I bring
I connect product strategy to systems that survive contact with reality.
0→1 product strategyDomain modelingInformation architectureSource contractsAI workflow designInterface behaviorQualification systemsHands-on implementation
“I can move from an ambiguous, high-stakes problem to a working product architecture — and make the tradeoffs legible to users, engineers, reviewers, and buyers.”
How I use AI
I use models to accelerate planning, search, summarization, and drafting inside deterministic product boundaries.
Identity stays explicit.
Evidence stays inspectable.
Unsupported work is withheld.
Human authority stays where it belongs.
The result: faster product development without outsourcing product judgment.
Product leadership for complex domains
Bring me the workflow where trust, speed, and messy source data all matter.
I can help turn it into a focused product strategy, a usable system, and a credible path to measured value.
Jay Sethi · jsethi.com
eMonograph + Pharma Data Pipeline · Case study, white paper, and portfolio · Jay Sethi · October 2026