Build vs. Buy — UT Southwestern PDF Remediation
Rough order-of-magnitude estimate of what it would cost UT Southwestern (an academic medical center) to build an in-house system that duplicates Gevell PDF, versus buying from us at volume. Prepared to support a quote for a customer wanting hundreds of thousands of pages/year of PDF remediation.
Status: rough estimate. Two inputs drive everything and should be confirmed with the customer: (1) actual annual page volume — modeled here at a 300,000 pages/year midpoint; (2) UTSW loaded software-engineer cost — modeled at ~$180k/year (medical-center base + ~1.4x benefits/overhead). Prior numbers updated from
docs/COMPETITIVE-ANALYSIS.md(market / manual-remediation rates) anddocs/admin/billing/skill-cost-model.md(our optimized per-page AI cost). Pipeline scope fromdocs/CONVERSION-PIPELINE.md.
What “duplicating our work” means
UTSW would not be building a PDF tagger — they’d be rebuilding a full pipeline. From our own architecture, the replaceable scope is:
- PDF ingestion, complexity classification, routing across processing paths
- Multi-pass vision conversion with screenshot-comparison refinement — the hard, defensible core
- MathML for printed and handwritten equations (Mathpix-class capability)
- Table extraction + degenerate-table re-extraction
- WCAG validation engine (45+ rules) + axe-core browser audit + auto-remediation loop
- PDF/UA structure tagging (Grackle-equivalent) if they also want tagged PDF
- Chunked async pipeline for large docs: scheduler, retries, orphan recovery, cost tracking
- Hosting, storage, auth, dashboards, human-review tooling
Our competitive analysis explicitly calls the vision-refinement + integrated-validation combination “architecturally difficult to replicate.” For an in-house academic team, this is a multi-year effort with real risk it never reaches parity on math/layout fidelity.
One-time build cost
Assuming ~$180k/yr loaded engineer cost:
| Item | Estimate |
|---|---|
| Core engineering team ~6 FTE × 12–18 mo | $1.1M – $1.6M |
| Accessibility SME (1 FTE) | ~$140k |
| PM + QA | ~$300k |
| One-time build to production quality | ~$1.5M – $3M |
Plus 12–24 months before it converts anything at compliance quality.
Frontier-model churn is a live tax: e.g., Google’s retirement of
gemini-2.5-flash broke one of our extraction paths overnight — an in-house
team absorbs that maintenance permanently.
Ongoing annual cost (at ~300k pages/year)
| Item | Estimate/yr |
|---|---|
| Maintenance/eng team (3–4 FTE — model updates, breakage, features) | $600k – $900k |
| AI inference + OCR/Mathpix + compute (unoptimized in-house; we run ~$0.01–$0.034/pg optimized, they’d run higher) | $50k – $150k |
| Human review on complex pages (~5% × ~15 min ≈ 2 FTE) | $150k – $350k |
| Ongoing all-in | ~$800k – $1.4M / year |
Effective annualized cost of building (3-year amortization of build + ongoing): roughly $1.5M – $2M / year, front-loaded and carrying execution risk.
Vs. buying from us
At this volume, per-page pricing drops well below the $1 standard rate:
| Scenario | 300k pages/yr |
|---|---|
| Buy @ $1/page (standard) | $300k |
| Buy @ ~$0.30–$0.50/page (volume) | $90k – $150k |
| Manual remediation @ $3–$25/page (the non-software fallback) | $900k – $7.5M |
| Build in-house (annualized, all-in) | $1.5M – $2M |
Bottom line for the quote
- Building in-house is roughly 5–15x more expensive per year than buying, plus a $1.5M–$3M upfront outlay and a 1–2 year delay during which zero pages get remediated — against a fixed ADA Title II deadline (April 26, 2027 for populations ≥50,000).
- The manual-services fallback is even worse: $0.9M–$7.5M/year at 300k pages.
- Even if UTSW can build it, they’d spend >$1M/year to reproduce something that costs them ~$100–150k/year to buy — while taking on model-maintenance risk and missing the compliance window.
Sensitivity
The two swing variables:
- Page volume. All ongoing/buy numbers scale with volume. At 200k pages the buy case is ~$60k–$100k (volume); at 500k it’s ~$150k–$250k. Build labor is largely fixed, so higher volume only widens the gap in favor of buying.
- Loaded salary. At $150k/FTE the build drops ~15%; at $220k it rises ~20%. Neither changes the conclusion.