Ayo Kadri
AI Systems Engineer
I'm an AI Systems Engineer — not just an ML practitioner. There's a difference: I don't hand off a model and call it done. I own the full stack: data pipelines, model training, serving infrastructure, monitoring, and the product layer that makes it useful to real people. MSc in Artificial Intelligence (University of Wolverhampton), and 6+ years building software and AI systems across healthcare, fintech, education, and AI products.
At Eyoto Group, I built the EyeVision AI Platform from the ground up — a production healthcare computer vision system that processes clinical ophthalmic imagery, automates annotation workflows, and deploys via Azure ML with full regression testing. At Your Study Path, I shipped LLM pipelines that moved the retention needle by 22% and cut deployment time by 60%. I'm now building CogFlow — an AI learning OS for professionals, powered by LangGraph multi-agent pipelines and pgvector.
My approach: architecture decisions matter more than model accuracy. A 94% AUC-ROC score in a system that breaks under load, can't be monitored, or can't be explained to stakeholders is not a success. I build AI systems that are scalable, reproducible, observable, and — most importantly — actually useful.
Google Cloud Certifications
Production systems deployed. Real metrics. Real domains.
Eyoto Group Ltd
Your Study Path Limited
Ascalon Integrated
StreetRates Limited
Education
University of Wolverhampton
AI Learning Operating System for Professionals
Problem
Professionals lack AI-native, adaptive learning tools. Most study tools are passive, generic, and disconnected from individual knowledge gaps. Learning should be as intelligent as the problems professionals are trying to solve.
Solution
A production-grade AI learning OS — not a chatbot wrapper, but a full system with multi-agent tutoring, RAG-powered curriculum retrieval, adaptive learning paths, memory persistence, and mastery analytics.
System Architecture
Multi-Agent Tutoring
LangGraph orchestrated agents
RAG Retrieval
pgvector + PostgreSQL
Adaptive Learning
Personalised study plans
Analytics Engine
Mastery & progress tracking
LangGraph
Multi-agent
pgvector
Vector store
Adaptive
Learning paths
Persistent
Memory system
Production AI platforms and shipped ML applications — each solving a real domain problem.
AI Learning Operating System for Professionals
Problem
Professionals lack AI-native, adaptive learning tools. Most study tools are passive, generic, and disconnected from individual knowledge gaps.
Healthcare Computer Vision Platform for Automated Annotation & Iris Detection
Problem
Ophthalmic software teams needed accurate, real-time iris/pupil detection and a scalable annotation infrastructure — not just a point model, but a complete AI system.
Credit underwriting teams needed interpretable, auditable risk scores — not black-box ML outputs.
SaaS businesses lacked actionable, real-time churn risk scores for customer success teams.
Full architecture breakdowns — problem, solution, challenges, and what I'd do differently.
Healthcare Computer Vision — Full Platform
Computer Vision / Healthcare10K+
Images Processed
70%
Annotation Effort Saved
Azure ML
Deployment Platform
Problem
Ophthalmic medical software needed real-time iris and pupil detection from clinical camera feeds — but the real challenge was scale: manual annotation of thousands of images was a bottleneck, and deployment to a safety-critical production environment required reproducibility and regression testing that didn't exist.
Solution
Built the EyeVision AI Platform end-to-end: automated CVAT annotation workflows with quality verification, YOLOv8 fine-tuning on 10K+ ophthalmic images, a streaming OpenCV inference pipeline, Azure ML for experiment tracking and deployment, and a regression testing suite for production releases.
Architecture
Challenges
Lighting variation across clinical cameras caused significant performance degradation on edge cases. Solved with targeted augmentation (albumentations) and domain-specific preprocessing. The annotation quality problem was harder than the model problem — inconsistent labelling was the real accuracy bottleneck.
Lessons Learned
Data quality infrastructure (annotation workflows, quality gates) should be the first investment in any computer vision project — not the model. I'd build the automated annotation pipeline before touching model architecture next time.
Real-Time Financial Risk Detection
MLOps / Fintech92%
Accuracy
25%
Fewer False Positives
<5ms
Inference Latency
Problem
A financial platform needed sub-10ms fraud classification at scale — before payment authorisation completed. Batch approaches were too slow, rule-based systems had unacceptable false positive rates triggering customer friction, and black-box model decisions couldn't be explained to compliance teams.
Solution
Built a fraud detection platform — not just a model. An ensemble classifier (XGBoost + shallow NN) backed by a Redis feature store for sub-millisecond feature retrieval. SHAP-based explainability for compliance. FastAPI microservice consuming Kafka events. MLflow for production monitoring and drift detection.
Architecture
Challenges
Feature freshness was the hardest problem. The model hit 92% in staging with batch features but degraded under real-time conditions where features were stale. Rebuilding around a Redis feature store with TTL management — before touching inference — was the critical architectural fix.
Lessons Learned
Build the feature store first. Every subsequent decision — model architecture, serving latency, monitoring — becomes simpler once feature consistency is solved. I'd also instrument for data drift from day one: by the time you notice it in production, the damage is already done.
The full stack — from raw data to deployed model to monitored API.
Google Cloud Certifications
Architecture choices, technical deep dives, and what I wish I'd known when building production AI systems.
LangChain got us to prototype fast. LangGraph is how we're shipping to production. Here's the architectural decision: what statefulness in multi-agent systems actually costs, why managed vector DBs add latency you don't need, and how pgvector inside PostgreSQL solved three problems at once.
The bottleneck wasn't the model — it was getting data into the model. This is the full story of building an automated CVAT annotation workflow for 10,000+ ophthalmic images: the edge cases that broke our quality checks, the Azure ML orchestration decisions, and what we'd do differently now.
Our ensemble model hit 92% accuracy in staging and nearly failed in production — not because the model was wrong, but because feature freshness was inconsistent. This is the architecture decision I'd make first next time: why we built a Redis feature store before touching inference, and how it changed everything.
Follow the build progress
I share architecture decisions, lessons learned, and behind-the-scenes engineering on LinkedIn.
Currently available for the following roles
Open to senior AI engineering roles, technical contracts, consulting, and collaboration on production AI systems.
I read every message. Usually reply within 24 hours.