★ Google GenAI Accelerator AwardAga Khan UniversityAWSAzureGCPEleutherAILF AI & DataLJMU · EPCC
I lead the data infrastructure, AIOps, and cloud engineering function at Aga Khan University, with 12 years building and running production data platforms across AWS, Azure, and GCP for distributed teams in Kenya and Pakistan. My MedGemma-based clinical decision support work, grounded in Kenyan Ministry of Health guidelines, was recognized through the Google GenAI Accelerator Award.
My platform work sits where data engineering meets AI — and increasingly, where AI meets evaluation. When our clinical RAG evaluation started rewarding the wrong behaviour, I built a deterministic scoring harness that exposed a failure mode standard accuracy had hidden across four releases, then published GC-Accuracy — the evaluation metric born from that harness, with an open benchmark. I also contributed the sole fix for a silent bug in EleutherAI’s lm-evaluation-harness that had left 35 African-language benchmark configs measuring nothing.
Taking a MedGemma-based assistant from prototype to daily use by doctors at Aga Khan University Hospital — and earning the Google GenAI Accelerator Award along the way.
The lakehouse, streaming, and governance foundation behind AKU's research and hospital analytics across Kenya and Pakistan — 99.9%+ uptime with p99 freshness SLOs.
Building the data platform for e-commerce across Kenya and Uganda: multi-AZ Postgres replication, 15-minute freshness SLOs, and a recommendation engine that cut referral payouts 60%.
I prove AI systems actually work, get them through governance into production, and build the data platforms that keep them there. Three ways to engage — each one flows naturally into the next.
Diagnose
AI evaluation audit & evaluation infrastructure
Your AI passes its evals — but are your evals telling the truth? I audit evaluation harnesses for the failure most teams never see: a metric that quietly rewards the wrong behaviour. In our own clinical RAG system, one such metric hid a failure mode across four releases; the deterministic, judge-free scoring approach I built to catch it became GC-Accuracy, an open evaluation metric and benchmark. I also found and fixed the silent bug that left 35 African-language benchmark configs in EleutherAI’s lm-evaluation-harness measuring nothing.
Fixed-scope audit, two to four weeks: I read your harness and metrics, reproduce what they actually optimise, and deliver written findings plus a corrected evaluation suite with regression CI your team keeps.
Deploy
Regulated AI — from prototype to governed production
Most GenAI prototypes stall where governance begins. I take AI systems through that wall: dosing-style safety gates, expert escalation paths, audit-ready traces, de-identification pipelines, and data-sovereignty architectures that legal, risk, and clinical teams can sign off. Afya Gemma, the MedGemma-based clinical decision support system I took from prototype to production, operates daily for doctors at Aga Khan University Hospital and won the Google GenAI Accelerator Award. Healthcare is the deepest reference; the same governance machinery transfers to fintech and any compliance-heavy sector.
★ Google GenAI Accelerator AwardIn daily clinical useNIH-funded consortium infrastructureHuman-in-the-loop by design
Typical engagement
Two to six months from working prototype to governed production: governance design with your risk owners, safety and evaluation gates wired into CI, then staged rollout with monitored escalation paths.
Sustain
Data & AI platform build-out — with fractional leadership
Both engagements above eventually meet the same wall: a data foundation that cannot support what the business wants. I build that foundation and the team that owns it — I have done it three times, and left it running each time. At Copia the platform sustained 250,000 daily transactions with 99.95% uptime and a recommendation engine that cut referral payouts 60%; at Zubale a re-architecture cut platform cost by roughly 25 percent; at Aga Khan University the platform holds 99.9%+ uptime across a three-country research consortium. Available as a build engagement or an ongoing fractional Head of Data & AI retainer.
★ 250K transactions/day sustained99.9%+ uptime SLAs−25% platform costTeams built from zero, three times
Typical engagement
Platform build-outs run three to nine months with your engineers embedded from day one; fractional leadership runs as a monthly retainer covering architecture, hiring, standards, and delivery cadence.
Also on the menu
Streaming & event platforms
MLOps & feature stores
LLM cost & latency engineering
Vendor & platform assessment
HPC & GPU optimization
BI & decision infrastructure
Technical due diligence
Team training & workshops
These sell through the three engagements above, not beside them — if one is the actual problem, we start there.
Field notes on data platforms, streaming, and applied AI in production.
These pieces are published on LinkedIn; future posts may also appear here on-site under /blog/.
Projects
Case studies of production systems I have led, and a selection of open-source work across data infrastructure, healthcare AI, MLOps, and applied AI. The full repository list lives on GitHub.
250K daily transactions through CDC pipelines, multi-AZ Postgres replication with <10s failover, and a recommendation engine that cut referral payouts by 60%.
GC-Accuracy: an open evaluation metric and benchmark crediting honest abstention in RAG systems. One retrieval fix read as a 4.5-point regression under standard accuracy — and an 8.5-point gain under GC-Accuracy.
MSc Dissertation: AI-Driven Workload and Energy Optimization for Exascale Scientific Computing.
★ 1
Ventures, Tools & Products
MvuukaFounder
Cross-border trade and payments platform for African informal traders (Delaware C-Corp). LangGraph agent architecture, USSD and WhatsApp interfaces, PAPSS escrow, HS-code classification, and duty and price estimation — built solo.
Manager, Data Infrastructure, AIOps & Cloud Engineering
~ Own data and AI platform delivery for AKU Global Data & Innovation, reporting to the Chief Data Officer. Lead distributed engineers across Kenya and Pakistan — hired the data engineering and data science team across both geographies and mentored two engineers into technical leads. Set SLAs for data uptime, quality, and freshness across the AKU Hospital data platform, the NIH-funded Uzima-DS Consortium (with the University of Michigan and KEMRI), and internal research teams.
~ Took Afya Gemma from prototype to governed daily production — dosing safety gate, safety-regression CI, and clinician escalation paths. The MedGemma-based clinical decision support system now operates daily for resident and intern doctors at AKU Hospital (Google GenAI Accelerator Award). Designed the two-stage retrieval architecture (Gemini 2.5 Flash classifier before MedGemma generation) on a ChromaDB vector store.
~ Built model evaluation infrastructure — a deterministic, judge-free scorer and safety-regression harness tracking leakage, fabrication, and drift — that exposed a silent failure mode standard metrics had missed across four releases, later formalised into the published GC-Accuracy metric.
~ Architected and scaled real-time and batch data platforms supporting research, analytics, and clinical operations across Kenya, Pakistan, and external consortium partners. Designed the secure Conversational RAG platform on Azure OpenAI, PgVector, and Azure AI Search with full data sovereignty.
~ Started the early architecture and prototyping of Afya Gemma.
~ Built foundational data pipelines and platform components. Built the university's first enterprise clinical data repository (records 2008–present). Designed multi-country ingestion pipelines including real-time Fitbit streaming for 615 healthcare workers in Kenya.
Kafka · Flink · Spark · Python · Azure Data Factory · Delta Lake · Airflow
Feb 2020– Aug 2022
Copia Global — Nairobi, Kenya
Senior Data Warehouse & Data Platform Manager
~ Architected the enterprise AWS data platform — Kafka, Spark, Glue, S3, Redshift, and Step Functions with CDC, monitoring, and alerting — sustaining 250,000 daily transactions for e-commerce across Kenya and Uganda. Postgres logical and physical replication across regions separated analytical from transactional workloads and eliminated analytics-driven downtime.
~ Built a LightGBM LambdaMART recommendation system on SageMaker behind live real-time APIs that cut referral payouts by 60%. Built the data function from scratch and left it owned by others: trained the team, wrote the engineering standards, and established governance adopted across four departments.
~ Re-architected the real-time event delivery platform — Kafka, Spark Structured Streaming, and ClickHouse on GKE with partitioning, backpressure handling, and spot-instance migration — cutting platform cost by roughly 25% — approximately USD 20,000 annually.
~ Reduced remediation backfills 40% with Avro schema contracts, automated data validation, and Prometheus/Grafana monitoring against freshness and accuracy targets; templated Airflow DAGs and CI/CD cut mean lead time for change by 30%.
~ Delivered ML models, dashboards, and customer 360 data marts. Directed analytics operations and partnered with architects to define BI strategy.
Apache Airflow · Azure Data Factory · Power BI · Tableau · SQL · Python · scikit-learn
Jan 2018– Apr 2019
Cropnuts — Nairobi
Data Analytics and Support Engineer (Digital Ocean)
~ Built the ML calibration model for spectral soil analysis that expanded laboratory testing access across Kenya, Senegal, and Nigeria, with engineering partners in the US and Germany.
Jan 2017– Dec 2017
SkyTOP Technologies — Nairobi
Software Developer
~ Designed custom software solutions to support BI adoption. Developed BI systems for SMEs.
Jan 2016– Dec 2016
Kenya National Examinations Council (KNEC) — Nairobi
Data Systems Analyst (Contract)
~ Supported secure data systems for national exam administration. Reduced processing time by 30%.
Jan 2015– Dec 2016
Redbric Consultancy — Nairobi
Assistant Research Officer
~ Research design and data collection across socio-economic development projects for NGOs.
Jan 2014– Dec 2015
Ahero Sub County Hospital — Kisumu
Health Information Systems Assistant
~ Managed patient records, extracted key statistics, digitized health records.
Education
2024–2026
Liverpool John Moores University — United Kingdom
MSc, Computer Science
~ Dissertation: RL and Bayesian Optimization for Workload Scheduling on HPC and Exascale Infrastructure — 12.3% energy reduction over first-come-first-served scheduling.
2010–2014
University of Kabianga — Kenya
BSc, Applied Statistics with Computing
Certifications
2025
Google Cloud — AI Infrastructure series: AI Hypercomputer, Cloud TPUs, Cloud GPUs
2025
HLRS, University of Stuttgart — Advanced Parallel Programming with MPI & OpenMP · GPU Programming Using CUDA · Unlocking HPC Performance with AMD Tools and GPUs
2025
BlueDot Impact — Technical AI Safety
2025
O’Reilly & Particular Software — Reinforcement Learning with LLMs · Designing Distributed Systems · Distributed Systems Design Fundamentals
2024–25
Linux Foundation & Tetrate — Kubernetes & Containers Fundamentals · Helm · Prometheus · Fluentd & Fluent Bit · Service Mesh · Istio Fundamentals
2024
NVIDIA & Red Hat — AI Infrastructure and Operations Fundamentals · Cloud-Native Development with OpenShift and Kubernetes
