ShelfSense

• The 98/100 feasibility score is directionally right on 'can it work' but conflates that with 'is it hard/defensible.' Retail shelf recognition is a solved-enough problem: SKU-level classification and planogram matching run on off-the-shelf detectors (YOLO-class / vision transformers) at 90-95%+ mAP on clean shelf imagery. Trax, Standard AI, and even Google Cloud Vision already do this — so the sci
• The genuine technical risk is NOT model accuracy in demos but robustness in the field: phone photos vary in angle, glare, occlusion, and store lighting; accuracy on messy real shelves (reflective packaging, near-identical SKU variants, private-label lookalikes) can drop 15-30 points vs. curated benchmarks. The moat, if any, is a proprietary labeled dataset of long-tail CPG SKUs plus a domain eval
• 'Agentic workflows / retrieval + tool orchestration' in the frontier tags is largely irrelevant hand-waving for this product — the core is supervised computer vision, not LLM agents. If the pitch leans on agentic-AI framing to justify the score, be skeptical: it signals narrative-chasing rather than the actual engineering (edge inference, SKU catalog management, planogram diff logic). Real value-a
• Traction is real but early and internally inconsistent: description says $26k MRR / 12 brands / 60k shelves; parsed plan metric shows ~$312k (annualized run-rate?). This discrepancy must be reconciled — a 12x gap between stated MRR and parsed revenue is a data-integrity flag even though it's not auto-detected as a RED FLAG.

• Thin-wrapper / commoditization: core CV is replicable in 6-12 months by a competent team on open models, and incumbents (Trax, Standard AI, ITsML) plus retailer-native tools already occupy the space. Without a proprietary SKU dataset moat, gross margins (~70%) compress as this becomes a feature, not a product.
• Field-robustness gap: no disclosed accuracy metrics on real-world uncurated photos. If false-positive out-of-stock/compliance alerts exceed ~5-10%, reps lose trust and churn — the '3x faster' claim is self-reported and unaudited. Demand a held-out field accuracy benchmark before conviction.
• Metric inconsistency ($26k MRR vs. $312k parsed revenue) plus no disclosed LTV/CAC or bottom-up TAM means neither stress test nor TAM triangulation could run — feasibility is de-risked but the business case is not.

• The 69/100 leans on a sector-reference ~70% gross margin and 35% capital intensity — reasonable for CV SaaS, but there is zero disclosed CAC, LTV, payback or churn, so the score reflects a template not this company. I'd mark it 'insufficient data' rather than 69; the number flatters a blank.
• Note the MRR discrepancy: description says $26k MRR (12 brands ≈ $2.2k/brand/mo, ~$312k ARR) while the model's parsed field reads $312k as MRR. If ARR is ~$312k, that is a real seed with land-and-expand potential; if MRR is truly $26k the team score's '$312k MRR' claim is wrong. Confirm before anything else.
• Unit-econ thesis hinges on seat/shelf expansion within each brand: 60k shelves/mo across 12 brands is shallow penetration of a national CPG's store base (a single big CPG audits 100k+ facings monthly). Net revenue retention >120% would confirm the land-and-expand story; flat or <100% NRR kills it given the low absolute MRR.
• Structural counter-argument: 'thin wrapper' risk (competitive headroom 41) directly threatens margins — if value sits at the model layer, inference costs compress the 70% GM and a well-capitalized incumbent (Trax, Snap/EDGE, ITL) reprices the category. Durable margin requires proprietary planogram-eval data + workflow lock-in, not just CV.

• No LTV/CAC/payback disclosed and stress test not run — a 3x-faster-visit ROI is a feature claim, not proven willingness-to-pay; if CAC exceeds ~12mo payback at $2.2k/brand ACV the model is uninvestable at seed.
• Established shelf-analytics incumbents (Trax, ITL/Shelfgram, Snap) already offer comparable CV compliance; a $3.5M seed buys little defensibility against their data moats and CPG relationships — churn to a bundled incumbent is the base-rate kill scenario.
• Concentration and TAM unproven: 12 brands, no bottom-up TAM triangulation. The ~$70B sector TAM is not ShelfSense's SAM — realistic SOM is field-audit software spend (low single-digit $B), and loss of 2-3 anchor brands would erase a third of revenue.

• The 57/100 market score is fair-to-generous. The '~$70B AI vertical SaaS TAM' is a category abstraction, not this company's SAM. Retail-execution/shelf-intelligence software (Trax, Repsly, Wiser, ISR/Pemberton-type tools) is realistically a $2-4B serviceable market; the addressable slice ShelfSense competes for is CPG field-rep audit tooling, not all vertical AI. I'd anchor to that narrower pool,
• Demand is inelastic where labor is the substitute: '3x faster store visits' converts directly to headcount/route economics for a CPG rep force costing $60-90k/rep fully loaded. That is a hard-dollar ROI pitch with genuine willingness-to-pay, which supports the ~70% gross-margin assumption. This is the strongest economic argument for the deal.
• Rents accrue to whoever owns the CPG data relationship and planogram integrations, not the CV model. $26k MRR across 12 brands = ~$2.2k/brand/mo — small, and pilot-shaped. Switching cost (the 59/100 moat) only becomes real once ShelfSense holds historical compliance data and is wired into brands' merchandising systems; at 60k shelves/mo that flywheel is unproven.
• Note the model's internal inconsistency: 'Team' cites $312k MRR while traction says $26k MRR — the parser appears to have annualized. Underwrite the $26k monthly figure; the composite's 64 team score is inflated by this artifact.

• Thin-wrapper / value-capture (structural, competitive headroom 41/100): commodity CV + LLM tooling means the defensible layer is integrations and proprietary shelf data, not the vision model. Trax and retail-media incumbents already own shelf-scanning at scale and can bundle it free into media/data deals — the disclosed-incumbent-substitution risk is real.
• Concentration and churn: 12 brands, ~$2.2k ARPA. Losing 2-3 pilots halves revenue. No LTV/CAC disclosed (stress test not run) and no bottom-up TAM (triangulation not run), so retention and true SAM are unverified — the two facts that most determine whether this is a business or a feature.
• Macro sensitivity: CPG marketing/field-execution budgets are cyclically cut first; a 3x-efficiency tool can be reframed by buyers as a headcount-reduction lever (shrinking seat count) rather than net-new spend, capping account expansion.

• Agree with 74/100, lean slightly higher on the merits: retail planogram/OOS detection is a benign use-case with no HIPAA/GLBA/hiring exposure. The main federal surface is FTC Act §5 on AI accuracy claims — '3x faster' and 'compliance' marketing must be substantiated or it becomes a deception theory. Cabin claims to measured pilot data.
• Data/privacy exposure is thinner than typical AI SaaS: shelf photos are property/product imagery, not PII, so CCPA/CPRA and the state patchwork are largely dormant unless reps' faces/bystanders are captured — require an incidental-face redaction policy and a B2B DPA passing store/CPG-owner image rights downstream. Colorado AI Act (effective 2026) targets 'consequential decisions' and does not reac
• IP posture is undocumented and this is the actual gap the memo must flag, not the regulatory score: no patents, trade-secret, or model-ownership disclosure. Given the composite's own 'thin wrapper' structural risk, verify ShelfSense owns its training data/annotations and eval harness rather than relying on a third-party vision API whose ToS could reclaim or throttle it. Value must accrue in propri
• Round mechanics are clean and standard: $3.5M seed via Reg D 506(b) (no general solicitation) with Form D in 15 days, or 506(c) with accredited verification if they market publicly — confirm every investor's Rule 501 status matches the exemption. Insist on 1x non-participating pref, broad-based weighted-average anti-dilution, pro-rata and information rights, and a board observer seat.

• Data-flow rights, not privacy law, are the sharp legal risk: if CPG customer contracts don't grant ShelfSense a license to retain/aggregate shelf images for model training, the proprietary-data moat is legally hollow and the 'thin wrapper' penalty (competitive headroom 41/100) hardens — the company owns software but not defensible data.
• Note the $26k MRR (description) vs. $312k MRR (parsed model input) contradiction — a ~12x discrepancy the model treated as fact. This is a diligence/representation-warranty issue: financial reps in the SPA/SAFE side letter must be verified, and any traction misstatement to accredited investors is itself §5 / anti-fraud exposure.
• FTC §5 substantiation: if '3x faster visits' or 'planogram compliance' underperforms in the field, a CPG customer or the FTC could frame marketing as deceptive; require claims to be tied to auditable eval-harness metrics and add a customer-facing accuracy disclaimer.