What it returns, what it costs all-in, exactly what we're buying, how it wires into your systems, and where it goes. No ranges, no open decisions.
The run-rate is $81K/year, and we're putting it on the front page. That's ~$3K of Azure plus a required managed-monitoring retainer (~$78K/yr) — a mandatory safety control for a classifier on a medical line, not an optional line item. Over five years, everything in, you spend ~$770K to return ~$5.25M in recovered margin. Net +$4.5M.
Starkey ships 1.4M units a year at $500–$1,600 of gross margin each. A three-person quality team spot-checks by eye and catches roughly 80%, worse late in a shift. A wrong-color shell, a wrong component, or a mislabeled unit that ships is rework, a return, or — for a mislabeled medical device — an audit or a recall. Automated vision inspects every unit at 95–99%+ with no fatigue. At the expected escape rate, the recovered margin is $1.05M a year.
Every verdict and image flows into your Microsoft Fabric data lake with a full 21 CFR Part 11 audit trail, validated against your Oracle serial-and-part master data, and becomes queryable in plain language across QMS, IMS, and line telemetry. "How many recalls this month, and what caused them?" answered in seconds, across every system — not a two-day BI request.
One inspection system, defect-escape prevention. Provable, underwritten, funds itself in months.
Quality-cost reduction, recall avoidance, and decision speed at platform maturity. Cost of quality runs 7–9% of sales in med-tech (McKinsey); capturing a few points on ~$1B is millions. Strategic tier — the reason to say yes to the direction, not just the pilot.
| Phase | What you get | Investment | Time |
|---|---|---|---|
| Phase 2 · Pilot | Live proof on your floor; business case validated against your data | $58,000 creditable | 3–5 wk |
| Phase 3 · Production | 3-station inspection, Azure MLOps, Oracle/QMS/IMS integration, audit trail | $285,000 | 4–6 mo |
| Phase 4 · Platform | Natural-language quality intelligence + agents over the Fabric lake | $37,000 | phased |
| Hardware (your CapEx) | $7,000 per station — 3 stations for production | $18,000 | one-time |
| Run-rate | Azure ~$3K/yr + required managed-monitoring retainer | $81,000/yr | ongoing |
| Pilot credits toward production | Net one-time build after credit | $343,000 |
Optional inspection add-ons, priced firm when you want them: white-label logo verification $26,000, metallic-pin protrusion measurement (R&D) $23,000. Every PRR fee maps line-by-line to committed engineer-hours and named roles — the full breakdown is in the Proposal tab.
| Item | Specified part | Qty | Cost |
|---|---|---|---|
| Camera | Basler ace 2 a2A2590-60ucBAS — 5 MP, color, USB 3.0 | 3 | $1,281 |
| Lens | Edmund Optics 25 mm C-Series #16528 | 3 | $1,122 |
| Dome light | Advanced Illumination DL2230 — diffuse, for true color | 2 | $2,154 |
| Ring light | Advanced Illumination RL121 — grazing, for serial OCR | 1 | $440 |
| Light controller | Strobe / hardware-trigger controller | 1 | $250 |
| Edge compute | Fanless industrial mini-PC — Intel i5, 16 GB, no GPU | 1 | $900 |
| Mounts + cabling | C-mount arms, USB3 locking cables, powered hub | set | $520 |
| Fixture + enclosure | 3D-printed fixture (you print) + light shroud | set | $333 |
| Committed station total | Three cameras: bottom (serial/OCR), top (color/assembly), oblique (logo) | $7,000 |
These are the exact parts — decided, not a range. Static parts in a fixed fixture mean rolling-shutter cameras and a CPU are sufficient; there is no need for a high-speed sensor or a GPU. Each additional production station is $5,500 (it shares the compute and controller).
Serial validation is what makes the verdict mean something — reading a serial isn't enough; matching it to the part, color, and logo it's supposed to be is what catches a traceability failure. It runs against your Oracle master data through the API Daniel exposes. All Azure consumption counts toward your Microsoft commitment (MACC) and is co-sell eligible.
Fund the pilot. Low cost, creditable, validated against your data, downside capped.
Agree the destination is a quality-intelligence platform, and this station is step one.