SAAIL Lab SAAIL Lab
PMLR Vol. 302 · DLI 2025/2026 Best Poster Award (Publication & Dataset Category) , Deep Learning Indaba 2025, Kigali 1+ Citations Published in PMLR Vol. 302 · Oral & Poster

Culturally Attuned and Resource-Aware Foundation Models for East African Agriculture: A Theoretical Framework and Research Agenda

Innocent Nyalala 1,2 , Nirav Bhatt 1
1School of Engineering and Science, Indian Institute of Technology Madras Zanzibar, Tanzania
2Wadhwani School of Data Science and AI (WSDSAI), Indian Institute of Technology Madras, Chennai, India
Deep Learning Indaba 2025 Research Track , Proceedings of Machine Learning Research (PMLR Vol. 302, pp. 1–11, 2026) (Jan 2026)
Culturally Attuned and Resource-Aware Foundation Models for East African Agriculture: A Theoretical Framework and Research Agenda

Summary: Best Poster presentation at Deep Learning Indaba 2025 (Kigali, Rwanda; published in PMLR Vol. 302): quantifies 5x-40x infrastructure disparities in East Africa and defines four architectural pillars plus Tiny/Lite/Edge deployment tiers.

Abstract

East African agriculture supports more than 175 million people yet faces mounting challenges from climate variability, compute scarcity, and information barriers. Current foundation models fail to address the region's computational constraints, linguistic diversity spanning over 200 languages, and distinct ecological knowledge systems. We present CARA-FM, a theoretical framework comprising four pillars: Community-Driven Data Architecture, Indigenous Knowledge Systems, Edge-First Model Design, and Participatory Governance. We define multi-dimensional evaluation metrics spanning technical, agronomic, and cultural criteria to guide foundation model research under severe resource constraints.

Keywords: CARA-FMfoundation modelsEast African agricultureresponsible AIindigenous knowledgePMLRDeep Learning Indaba

Key Findings & Contributions

PMLR 302
Published Proceedings
pp. 1-11 (DLI 2025 Oral)
4 Pillars
CARA-FM Architecture
Data, IKS, Edge, Governance
10x-40x
Infra Resource Gap
RAM & bandwidth reality
5M-50M
Edge Compression Tiers
Tiny / Lite / Edge specs
  1. Quantifies concrete resource mismatches between standard foundation models (8-32 GB RAM, 50+ Mbps) and East African field realities (1-3 GB smartphones, 1.2 Mbps connectivity, intermittent power).
  2. Formulates four interdependent design pillars: Community-Driven Data Architecture, Indigenous Knowledge Systems (IKS), Edge-First Model Design (5M-50M params), and Participatory Governance.
  3. Defines multi-dimensional evaluation criteria balancing technical accuracy with local linguistic coverage, agrarian equity, and energy constraints.

Key Empirical Figure

Best Poster Award recognition at Deep Learning Indaba 2025, Kigali, Rwanda (published in PMLR Vol. 302, 2026).
Best Poster Award recognition at Deep Learning Indaba 2025, Kigali, Rwanda (published in PMLR Vol. 302, 2026).

Benchmark Results

East African Edge Reality vs. Standard Foundation Model Assumptions

Resource DimensionStandard FM AssumptionEast African Field RealityGap Factor
Model Memory (RAM)8 - 32 GB1 - 3 GB10x
Network Bandwidth50+ Mbps1.2 Mbps40x
Power AvailabilityContinuous4 - 8 hours/day6x
Target Device Cost$500+< $1005x
Cloud GPU AccelerationRequiredUnavailable offlineInf

Proposed CARA-FM Edge-Cloud Deployment Tiers

TierParameter / Size BudgetTarget HardwareTarget Acc.Primary Use Case
CARA-Tiny5M params / 20 MBEntry feature & low-end phones80%Offline crop symptom screening
CARA-Lite15M params / 60 MBBudget smartphones (1-2 GB RAM)85%Multi-disease field triage
CARA-Edge50M params / 200 MBMid-range extension phones90%Full seasonal advisory

BibTeX

@inproceedings{nyalala2026carafm,
  title={Culturally Attuned and Resource-Aware Foundation Models for East African Agriculture: A Theoretical Framework and Research Agenda},
  author={Nyalala, Innocent and Bhatt, Nirav},
  booktitle={Deep Learning Indaba 2025 Research Track, Proceedings of Machine Learning Research (PMLR)},
  volume={302},
  pages={1--11},
  year={2026},
  url={https://proceedings.mlr.press/v302/nyalala26a.html}
}
APA Citation

Nyalala, I., & Bhatt, N. (2026). Culturally Attuned and Resource-Aware Foundation Models for East African Agriculture: A Theoretical Framework and Research Agenda. Proceedings of Machine Learning Research (PMLR), 302, 1–11. https://proceedings.mlr.press/v302/nyalala26a.html