PartSnap — nameplate-to-part scout for kitchen equipment techs

1. One-liner

Photograph a grease-caked equipment nameplate; get the exact OEM part and likely fault before you leave the kitchen.

2. Trend signal — why now?

The expensive moment in commercial foodservice equipment repair isn’t the fix — it’s the misidentification. A tech kneels behind a fryer in a 95°F line, the data plate is caked in carbonized grease or the stamped model field has worn smooth, and they guess. Guess wrong and you’ve burned a second truck roll: $150–300 of windshield time plus a restaurant with a dead fryer for another 2–5 days while the right part ships (d-tools truck-roll guide, 2026; Smart Service trip-charge data, 2026).

Three things converged in the last 12 months:

1. The ID problem is industry-acknowledged, not a hunch. Foodservice trade press states plainly that stamped model/serial plates “wear off in places and become illegible,” and that this is “a recognized problem in the commercial kitchen equipment industry” (Heritage Parts via FES Magazine). Parts distributors’ own answer is a web form where you type a clean serial in — which is exactly the input that doesn’t exist when the plate is unreadable (Parts Town Serial Number Lookup).
2. Parts-on-truck is the #1 lever on the industry’s headline KPI. First-time fix rate benchmarks at ~80% across field service; under 70% “puts the success of your organisation at risk,” and “parts readiness is the #1 driver of first-time fix rate” (Comparesoft FTFR; Smart Service).
3. Vision LLMs got good enough in 2025–26 to read a corroded plate at an angle. Modern multimodal OCR clears 98–99% page-level accuracy and handles “challenging scenarios,” and cheap inference made per-photo cost trivial (SnapCall OCR guide, 2025; BentoML VLM survey, 2026). Pre-2025 OCR choked on grease, glare, and stamped (not printed) characters.

Provenance:

• Signal 1 (Demand): Worn/illegible equipment nameplates are an industry-recognized problem; distributors’ lookup tools require a clean serial as input — Heritage Parts / FES Magazine + Parts Town — observed 2026-05-17
• Signal 2 (Economic): Wrong/missing part = $150–300 per second truck roll; first-time-fix is the industry’s headline KPI and “parts readiness is the #1 driver” — d-tools + Smart Service — observed 2026-05-17
• Signal 3 (Feasibility): Multimodal OCR/VLMs reached 98–99% accuracy on hard scenarios with cheap inference in 2025–26 — SnapCall + BentoML — observed 2026-05-17 Category: Tech-unlock

3. The opportunity

Parts Town, Allpoints, Heritage and the other foodservice distributors are enormous and competent — at selling parts. Their economic incentive is order volume, not de-risking your truck roll. Their flagship “innovation,” serial-number lookup, assumes the failure case away: it needs you to read and type a serial that, in the exact moment you need help, is illegible. There is no incumbent that takes a photo of a destroyed nameplate (or the broken assembly itself) and returns: decoded make/model/serial → exact OEM part number + aftermarket equivalents → the 2–3 most likely faults for that symptom on that model → whether that part is in this tech’s truck or the shop’s warehouse right now.

The horizontal FSM platforms (BuildOps, ServiceTitan, Housecall Pro) own scheduling and invoicing but treat parts as a generic line item — they have no foodservice parts taxonomy and no vision layer. The wedge is the 90 seconds between “I found the problem” and “I ordered the wrong thing.” Own that moment and you sit on top of every parts order in the shop.

4. Target market

• Primary customer: Owner or service manager at an independent commercial foodservice equipment service company in the US — the regional players with 5–40 field techs (think the hundreds of CFESA-member shops below Whaley/Tech-24 scale). Secondary user: the field tech, in-app.
• Why they buy: “My newer guys guess parts and I eat the return trip.” Tech shortage and an aging workforce mean shops are staffing with juniors who don’t yet have the model catalog in their heads (Total Food; PartsFE). Every avoided second trip is $150–300 they keep and a restaurant they don’t lose.
• Rough TAM reasoning: CFESA represents the organized core of US foodservice service companies; the long tail of independent shops runs into the low thousands. At a few thousand shops × $300–800/mo, the serviceable revenue is comfortably in the $10–40M ARR range — too small for ServiceTitan to chase a vertical feature, perfect for a bootstrapper.
• Why now for them: The labor mix shifted toward juniors and the AI to compensate for missing senior pattern-recognition just arrived. They feel the FTFR pain monthly and have no foodservice-specific tool aimed at it.

5. Product sketch (MVP)

• Snap-the-plate: photograph a nameplate (grease, glare, stamped, angled) → decoded manufacturer, model, serial, age/series.
• Snap-the-part: photograph the failed assembly when there’s no plate → candidate part matches with confidence and reference images.
• Exact OEM part number + listed aftermarket equivalents, with distributor deep-links (Parts Town / Allpoints / Heritage).
• Symptom → likely-fault shortlist: pick the reported symptom, get the 2–3 most common failing parts for that model.
• Truck/warehouse stock check: “you have this on the van” / “shop has 2, order for will-call.”
• One-tap parts order or a clean parts list pushed to the shop’s existing distributor account / FSM.
• Per-tech and per-shop history: every ID logged against the job for warranty and recall lookups later.

6. AI angle — what’s load-bearing

Remove the AI and there is no product. The entire wedge is a vision-language model reading a degraded, non-standard data plate that defeats classical OCR (stamped characters, carbon film, reflection, oblique angle) and then reasoning across messy, inconsistent OEM catalogs to resolve model → exact part — plus the symptom→fault inference that replaces the senior tech’s pattern memory. This is not OCR-as-feature; it’s OCR + catalog reasoning + diagnostic inference as the entire core loop. A non-AI version is just the serial-lookup web form that already exists and already fails on this exact input.

7. Localization angle (if any)

N/A — this is a US-first play. The wedge is US foodservice equipment OEMs, US distributor catalogs (Parts Town/Allpoints/Heritage), and the CFESA-organized service-company market. The same shape could later port to UK/EU and India HoReCa, but US is where the parts taxonomy, the distributor deep-links, and the willingness-to-pay are densest. Forcing a localization angle would weaken, not strengthen, the v1.

8. Business model — path to $1M–$5M ARR

• Pricing: Per-shop SaaS with a per-tech component — $99/mo base + $25/active tech/mo. A 12-tech shop ≈ $399/mo; a 30-tech shop ≈ $849/mo.
• ACV: ~$5,500/shop/year blended.
• Rough math to $1M ARR: ~180 shops at ~$5.5K ACV. CFESA’s organized membership alone makes 180 reachable.
• Rough math to $5M ARR: ~900 shops, or fewer shops plus a parts-distributor affiliate/referral cut on the order volume routed through PartSnap (distributors pay for qualified, correctly-specced orders — a real second revenue line once volume exists).
• Expansion path: seats grow as shops hire; usage tier on volume; warranty/recall lookup module; an OEM-paid “verified part” placement; eventually a managed parts-procurement upsell.

9. Go-to-market wedge — first 100 customers

• CFESA is a list, not a vibe. CFESA publishes industry standards and convenes the organized service-company market (cfesa.com/industry-standards). Build a 30-second teardown video using a real ugly nameplate, then directly work the member/exhibitor roster: personalized outreach to owners + a booth/demo at the CFESA annual conference where every buyer is in one room.
• Ride-along proof loop. Recruit 5 shops for a free 30-day pilot; instrument first-time-fix-rate before/after on equipment-ID jobs. A measured 3–6 pt FTFR lift is the entire sales deck for shops 6–100.
• Distributor reps as channel. Parts Town/Allpoints territory reps already sit in front of every shop and are measured on order accuracy — a co-marketed “snap it, we’ll spec it right” reduces their returns. Pilot a referral arrangement with one distributor’s regional team.
• Trade press + the foreman channel. FES Magazine “Service Insights” and the CFESA newsletter reach exactly these owners; a single case study with hard FTFR numbers placed there is high-signal in this small world.

10. Build complexity — justification

Medium. The hard parts are off-the-shelf: a frontier multimodal model handles plate-reading and symptom reasoning; mobile capture and a web console are standard stack. The genuine work is (a) assembling and continuously cleaning a cross-OEM model→part reference (scraped/licensed distributor catalogs + a feedback loop where techs confirm/correct matches — this is the moat-building grind) and (b) tuning the vision prompt against a corpus of deliberately awful real-world plate photos. A focused pair ships a credible v1 in ~3–4 months; the parts dataset deepens for years after.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Reading public part data and linking to distributors; respect catalog ToS, use licensed/partner feeds where required.
Ethical — no harm / dark patterns	✅	Reduces wasted trips and downtime; aligned with customer interest.
Market exists (evidence above)	✅	Industry-acknowledged ID problem; quantified second-trip cost; organized buyer body (CFESA).
1–5 person team can build this	✅	Off-the-shelf vision + standard stack; dataset is the effort, not infra.
Launchable with <$50K / ₹40L	✅	Inference + dataset assembly + a pilot tour. Comfortably sub-$50K to first revenue.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Felt multiple times/week per shop; directly costs $150–300/incident + lost customer. Not hair-on-fire daily for the owner personally, hence not 18+.
Demand evidence	15	12/15	Industry press explicitly names the illegible-plate problem and the FTFR/parts-readiness link; distributors built (inadequate) tooling. No direct verbatim “I’d pay” quotes pulled — keeps it at 12.
Build feasibility	15	11/15	Vision is solved; the cross-OEM parts dataset and prompt-tuning against bad photos are real 3–4 month work.
Distribution clarity	15	11/15	CFESA roster + conference + distributor reps is a concrete channel; conversion math still unproven.
Revenue mechanics	15	12/15	Pricing benchmarks against FSM seats; 180 shops to $1M is plausible. Distributor affiliate line is upside, not banked.
Time to first revenue	10	8/10	Pilot-to-paid in ~6–8 weeks; needs a 30-day proof period to show FTFR lift.
Defensibility	10	6/10	Soft moat: accumulating tech-confirmed model→part corrections + workflow lock-in. Copyable by a distributor, but they lack the incentive (they profit from order volume, not accuracy) — that misalignment is the real edge.
Total	100	76/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required — needs a builder comfortable with vision pipelines and either a foodservice-service-industry advisor or a founder who’s carried a service truck. Without domain depth the parts taxonomy will be wrong in ways customers instantly spot.

Key assumptions to validate (3–5)

1. Assumption: A frontier VLM reliably decodes genuinely degraded stamped nameplates (grease, glare, oblique) at field-usable accuracy. How to test: Collect 200 real bad-plate photos from 5 shops; measure model/serial extraction accuracy vs. ground truth. Target ≥85% usable.
2. Assumption: Shops attribute a measurable FTFR lift to the tool, not noise. How to test: Instrument before/after FTFR on equipment-ID jobs across 5 pilot shops for 30 days; require a ≥3 pt lift.
3. Assumption: Owners will pay ~$400–850/mo for it. How to test: 25 owner interviews + 5 paid pilots converting at the quoted price, not a discounted one.
4. Assumption: A cross-OEM model→part reference can be assembled legally and kept current. How to test: Audit distributor catalog ToS; secure one licensed/partner feed before committing to scrape-only.

Risk flags

1. Platform/data dependency: Reliance on distributor catalogs for part mapping — mitigate with a licensed feed and a tech-confirmation dataset that becomes proprietary over time.
2. Incumbent reaction: A distributor could bolt photo-ID onto their lookup. Mitigant: their incentive is order volume, not accuracy/returns reduction; speed + neutrality across distributors is the counter-position.
3. Accuracy trust: One confidently-wrong part match early in a pilot poisons word-of-mouth in a small, tight industry. Mitigant: always surface confidence + a “verify before ordering” affordance; never auto-order on low confidence.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder (vision/ML) + foodservice-service-industry advisor
Time to revenue:        6–8 weeks (after a 30-day pilot proof period)
Capital to launch:      $15–30K (inference, dataset assembly, CFESA pilot tour)
Top 3 assumptions to validate first:
  1. VLM accuracy on 200 real degraded nameplates ≥85% usable
  2. ≥3 pt first-time-fix-rate lift across 5 instrumented pilot shops in 30 days
  3. 5 shops convert to paid at full $400–850/mo, not a discount
Kill criteria:
  - Abandon if VLM model/serial extraction stays <70% usable on real bad plates after prompt+pipeline tuning
  - Abandon if pilot shops show no measurable FTFR lift after 30 instrumented days
  - Abandon if a major distributor ships free photo-ID with neutral multi-distributor results before your v1

15. Next step — 1-week validation sprint

• Day 1–2: Get 200 real degraded nameplate photos (call 5 shop owners, offer a $50 gift card per tech for a batch of ugly-plate pics with known model/serial). No synthetic data.
• Day 3–4: Run the photos through a frontier VLM with a tuned extraction prompt. Score model/serial accuracy against ground truth. Interview 10 of those owners on second-trip frequency and what they’d pay.
• Day 5: Go if (a) extraction ≥85% usable AND (b) ≥6 of 10 owners independently quote second-trip pain costing them real money monthly AND name a price ≥$300/mo. No-go otherwise.

Falsifiable: a hard accuracy number on real photos + a hard count of owners naming a price. Not “they seemed interested.”