Deciding on the premise that it will spoil — reading produce quality with AI and carrying it straight into routing
The National Science Foundation awarded roughly $1.21 million to BETAFELD to integrate AI quality grading of produce with prescriptive optimization guiding routing and utilization. The aim is to shift from static image classification to dynamic, decision-driven management of assets that degrade with time.
Grant overview (primary data)
- Award amount$1,207,205
- RecipientBETAFELD LLC (Massachusetts)
- ProgramSBIR Phase II
- Period2026-08-01 〜 2028-07-31
- FunderU.S. National Science Foundation (NSF) / NSF
Key points
- Integrating AI quality grading of produce with prescriptive optimization guiding routing and utilization ($1,207,205).
- The high-risk element is not grading itself but carrying grades into decisions within highly variable supply chains.
- The core contribution is a unified representation of defects that generalizes across categories, paired with a decision framework accounting for temporal decay and handling variability.
- It shifts from static image classification to dynamic, decision-driven management of perishable assets.
- Performance is validated through pilot deployments, measured by prediction accuracy, reduced shipment rejections and optimized utilization.
1Getting the appearance right is not enough
Grading produce from images has many precedents. What this project names as its high-risk element is not the grading but the part that carries a grade into routing and allocation decisions. Produce keeps degrading from the moment it is graded. Something judged good today is not guaranteed to be in the same state on arrival three days later. Treat grading as static classification and that temporal element drops out.
2From static classification to dynamic management
The right-hand column is what the project names as its core intellectual contribution. Rather than representing defects differently per category, it unifies the representation and pairs it with a decision framework that carries the passage of time. Coupling real-time AI outputs with decision frameworks, the description notes, is a known deep tech hurdle requiring profound cross-domain integration.
3A staged approach
The methodology proceeds in stages. First, curating diverse datasets across produce types and seasonal conditions to capture real-world edge cases. Then developing AI models using transfer learning and structured data augmentation to ensure robustness and prevent overfitting.
Those quality signals are then ingested by prescriptive optimization models factoring in storage conditions, transport constraints and operational trade-offs. Performance is validated through pilot deployments, measured by prediction accuracy, reduced shipment rejections and optimized product utilization.
4In the context of food waste
The project states that reducing rejected shipments and improving recovery of surplus produce expands access to lower-cost food beyond the financial gains, supporting a more resilient and enduring food system by ensuring usable produce reaches appropriate consumers. Of the NSF awards this site holds as of 2026-09-02, seven are SBIR Phase II, and companies appear among the 93 institutions. Amounts are the obligated amount as of the check date and may change.
Why it matters
That the hardest part is carrying model output into actual decisions applies wherever AI is embedded in operations. Designing for assets whose value changes with time also carries beyond produce into inventory management generally.
FAQ
Why is grading alone insufficient?
What is prescriptive optimization?
Sources (primary)
Source: NSF Award Search (U.S. National Science Foundation, public domain). Amounts are the obligated amount. For privacy, we do not handle principal investigator names.
- NSF Award (original, official)
- NSF Award ID: 2537269