How I think about AI operations
The model is only one part of the system. Good outcomes also depend on retrieval, memory, tool access, decision rules, evaluation, escalation and a person who understands the consequences of being wrong.
AI operations and product systems
I design and operate AI-enabled workflows.
My work is the layer between an AI model and a useful operational outcome: defining what the system should do, organising the context it needs, connecting tools, directing implementation, reviewing its decisions, and improving the workflow when reality exposes a weakness.
Profile
Practical AI judgment, developed through operating real systems.
I am strongest in roles where a company already has capable AI tools but needs someone to make them reliable, understandable and useful across everyday work.
What I bring
Product judgment, operational curiosity, clear requirements, strong constraint checking, factual skepticism, live incident handling, documentation and an instinct for turning recurring friction into a reusable workflow.
Selected work
Two AI systems I direct and operate.
They solve different problems, but both demonstrate the work I want to continue doing professionally.
A persistent AI operations workspace
RoseX lets me work with an AI agent through Telegram while preserving access to project context, files, browsers and operational tools. I shaped it around the problems that made ordinary chat interfaces unreliable for long-running work.
- Separated stable instructions, current operating context and durable memory.
- Added visible progress, same-turn steering, multi-file handling and restart recovery.
- Kept human authority explicit around interruption and production risk.
A real-time content decision system
The Content Engine tracks football fixtures, interprets changing provider data, prepares visual content and publishes subscriber updates. I define how it should behave when sources are incomplete, late or contradictory.
- Designed lifecycle rules, timing gates, refresh checks and operator escalation.
- Evaluate factual accuracy, timing, state progression, asset quality and publication behaviour.
- Use incidents as regression scenarios and shadow-test new decision logic before authority.
Working method
How I take an AI workflow from idea to daily use.
Understand the work
Start with the user, the operational environment and the cost of an incorrect or incomplete outcome.
Define the system
Specify the outcome, inputs, context, tools, authority boundaries, failure cases and visible acceptance criteria.
Direct implementation
Use AI development tools to translate the design into working software while reviewing trade-offs and protecting the intended behaviour.
Evaluate the execution
Inspect source data, intermediate state, logs and final output. A fluent answer or successful command is not enough.
Retain the learning
Turn failures and operator feedback into improved rules, tests, monitoring, documentation and memory.
Experience
Product operations and technology programme delivery.
2024–2026
Screentime SportsProduct & AI Operations Lead
Own requirements, AI workflow direction, live operations, QA, customer access, payment operations and incident response for an independent real-time sports information service supporting more than 300 paying subscribers.
- Directed interconnected customer, administrative and content workflows used in live production.
- Converted recurring failures into monitoring, recovery controls, safer operator tools and documented procedures.
- Prioritised changes by customer impact, evidence and operational risk.
2023–2024
Imisi 3D / ERDITechnology Programmes & Lab Facilitator
Supported technology programme delivery within an established African XR creation lab and UNICEF Innovation Fund graduate.
- Facilitated XR masterclasses, demonstrations and adoption sessions for corporate groups, institutional guests, schools, startup visitors and public audiences.
- Translated unfamiliar immersive technology into practical guidance and resolved equipment and session-delivery issues.
- Supported engagements involving Air Peace, the Federal Ministry of Education, CcHUB Creative Economy Practice and Omniverse Africa.
2023–2024
ARVR AfricaCommunity Programme Support
Supported programme communication, participant engagement and collaborative learning within a continent-wide XR creator and professional community.
Programme work
VR4SchoolsImmersive Learning Delivery
Guided secondary-school students through curriculum-aligned AR/VR learning experiences, adapted technical explanations to participant needs and collected structured post-session feedback.
Tools and environments
Tools I use to move work forward.
The value is not the length of the tool list. It is knowing which instrument fits the task and how to verify what it produced.
ClaudeOpenAI CodexChatGPTTelegramBrowser automationMCP connectorsAPIsGitHubGoogle WorkspaceMicrosoft OfficeOperational dashboards
Technical preparation
Concepts reviewed for AI operations work.
I reviewed these materials to connect formal terminology with patterns I already encounter while operating agentic workflows.
Agent evaluations and traces
DeepLearning.AI · execution-path analysis, observability and evaluator design.
GraphRAG
Microsoft walkthrough · relationship-aware retrieval across connected knowledge.
Hybrid retrieval
English walkthrough · combining semantic and keyword retrieval.
Auto-research loops
Andrej Karpathy discussion · iterative research, experimentation and feedback.