Reliable AI evaluation
How benchmark design, evaluator choice, prompt framing, and language coverage affect conclusions about model performance, particularly for African languages.
RELIABLE AI · SMALLHOLDER AGRICULTURE · INCLUSIVE SYSTEMS
I’m Elvis Gyasi Owusu, a Ghanaian technology founder, software engineer, and researcher working across agricultural infrastructure, accessibility, and AI evaluation. I build offline-capable production systems and study how evaluation design shapes conclusions about AI.
My work connects trustworthy evaluation with the design of useful digital systems for African languages, smallholder agriculture, accessibility, and low-connectivity environments.
How benchmark design, evaluator choice, prompt framing, and language coverage affect conclusions about model performance, particularly for African languages.
How to design and evaluate useful tools for farmers and service providers when connectivity, language coverage, and data are limited.
How offline-capable architecture, accessibility, and technical reliability affect whether AI-enabled services work in everyday use.
Current work measures how evaluator family, assertive framing, verification requests, and small wording changes can alter the conclusions drawn about a language model.
My systems work examines how offline-capable architecture, accessibility, and technical reliability determine whether AI-enabled services remain useful in everyday Ghanaian contexts.
Two 2026 arXiv preprints in language-model evaluation and African-language AI, with manuscripts under review at TACL and ACL Rolling Review.
Manuscript under review at TACL · submitted September 2026
Examines how evaluator model family can systematically change pairwise LLM judgments. Across 9,312 judgments, the study introduces a corrected estimator that separates family-conditioned preference from candidate quality.
Manuscript under review at ACL Rolling Review · October 2026 cycle · Submission 218
Preprint title: AfriSyCo: Measuring Assertive Framing, Verification, and Wording Sensitivity Around African-Language Content
Studies answer switching across six African languages and seven open-weight checkpoints. It shows that assertive framing, verification requests, and small wording changes materially affect measured model behaviour.
The roles that best explain the through-line between company building, engineering, teaching, and research.
Company strategy, product direction, technical architecture, partnerships, and field operations for digital agricultural infrastructure.
Building an AI-powered platform for coordinated content, publishing, engagement, analytics, and growth workflows.
Led accessibility-focused product engineering across web platforms, browser extensions, reusable widgets, and shared AI services.
Supported undergraduate teaching and mentored students in programming, software engineering, computer networks, and research practice.
A focused portfolio of platforms I have founded, led, researched, or engineered.

Offline-first agricultural service infrastructure connecting farmers, mechanisation providers, and field teams through mobile applications, USSD, SMS, and Mobile Money.
21,759 registered farmers · 3,554 providers · 9 regions · 16 districts · 6,273 service requestsView project
A multi-tenant AI orchestration platform for content planning, generation, approvals, publishing, engagement, analytics, and lead qualification across social platforms.
13 specialised AI agents · multi-platform publishing · real-time workflowsView project
A public consulting website and companion operations platform for content, enquiries, email workflows, image management, secure administration, and privacy-conscious analytics.
Marketing site + secure admin · resilient content fallback · first-party analyticsView project
An accessibility extension and embeddable web widget supporting speech, translation, summarisation, contrast, and interface adaptation.
Deployed across 100+ websites · 500+ weekly npm downloadsView project
A paired mobile safety system combining SOS reporting, live location, responder coordination, secure messaging, and anomaly detection over movement trajectories.
First-class final-year research · paired Flutter apps · DBSCAN anomaly detectionView project
Student, administrator, and super-administrator platforms for accommodation discovery, booking, payments, communication, and operational management.
4,000+ students · 4+ hostels · three connected portalsView projectFinal-year research: CampSafe, an AI-powered campus safety platform using unsupervised anomaly detection over location trajectories for automated emergency escalation.
View degree certificateHall President, 2018—2019.
View WASSCE certificateSelected completed credentials with direct verification links.
Professional certificate covering project initiation, planning, execution, and agile delivery. The Credly credential is endorsed by FIBAA and evaluated by the American Council on Education for college-credit recommendations.
Verify credentialCompleted specialization in prompt design, AI-assisted analysis, creative collaboration, and workflow optimization.
Verify credentialApplied Jira automation for repeatable project workflows, issue transitions, and delivery processes.
Verify credentialSpecialization authorized by IBM covering React front-ends, modular server-side applications with Node.js and Express, NoSQL data in MongoDB, and RESTful API design.
Verify credentialBuilding scalable backend services integrated with AI using Node.js and MongoDB.
Verify credentialCrossBoundary Advisory · Mastercard Foundation initiative
BUILD · RESEARCH · COLLABORATE