DocNexus.ai · case study
Market sizing engine
National market size and share for radiopharmaceuticals, from a partial claims panel, with every number carrying its source and its band.
- 49
- products sized
- 29
- competitive sets
- 1.72–1.99×
- lens divergence (2024)
- 5 of 5
- validation gates
Deterministic estimation, explicitly not forecasting: national volume, patients and value for 49 products across 29 clinical competitive sets, projected from a partial panel and anchored to public sources. A catalog scale-up for 347 drugs completed its first four of twelve planned phases.
Methods and relative results only. Client figures, internal names and scale are not published.
Projection that admits what it knows
A flat coverage gross-up is biased whenever a drug’s mix differs from the panel’s. It was replaced with a stratified per-cell projection (drug × payer × site × geography), measured against manufacturer disclosures where one exists. Measured panel coverage for the anchored drugs is 41–64%.
Coverage comes in three honesty tiers: measured, imputed from the competitive set, or assumed. Share is robust to the assumption; absolute counts are labelled directional.
Two value lenses, never merged
Manufacturer net revenue (from SEC filings, shown only for brands that disclose it, never fabricated) and payer-allowed spend (what payers actually reimburse) differ by 1.72–1.99× by design: provider margin plus gross-to-net. A proposal to collapse them into one number with a fixed ratio was rejected, because the ratio came from a handful of branded drugs and would have invented net values for generics.
Volume ships as a band, not a false point, because these drugs are not in the federal average-sales-price file and net price is an estimated range. An audit forced the band to be relabelled from “triangulated” to price sensitivity, which is what it is.
Architecture worth drawing
An offline DuckDB and ClickHouse pipeline writes a small committed JSON file; a statically generated Next.js app renders it. The deployed dashboard has no database or query engine at runtime and needs no secrets. Same inputs, byte-identical outputs.
The catalog scale-up adds a Postgres, FastAPI and job-queue backend with 17 US government-data connectors. A publish rule auto-publishes only from allowlisted, deterministically parsed government sources (19 of them) and routes everything else (17) to human review.
A better coverage method, approved but not built
Two independent designs were merged into a new coverage method. It counts clinical events, not claim lines: lines over-count events by 1.6–2.1×. It carries each drug’s own history forward (6.6–8.1% median error in backtests). It rejects the office physician file as a total anchor (46.5–53.3% error) and capture-recapture, which overstated capture by 1.5–2.4×.
Decisions
- The code → drug → modality crosswalk was sourced by two independent passes over 55 codes, then re-verified by an adversarial audit.
- Validation gates (band ordering, no negatives, coverage reconciliation, decomposition sums) pass 5 of 5.
- A fresh-agent audit raised 25 findings; 23 were confirmed and fixed with regression tests, and the other two were rejected with reasons.
Where it falls short
- The event-count coverage method was designed, backtested and approved, but not built, so the shipped dashboard still measures coverage on claim lines. Built, it would lower measured coverage by 29–38%.
- The catalog scale-up stopped after its first four phases: US sources only, and not released.
- Absolute volumes for drugs without a disclosed anchor are directional.