Data Scientist & AI Engineer · Causal ML · MLOps
3+ years building intelligent systems — from real-time Kafka pipelines and PyTorch deep learning models to multi-agent LangGraph systems with Claude & OpenAI, causal ML platforms, and live production deployments with paying users.
I'm a Data Scientist and AI Engineer with 3+ years building production-grade ML platforms, agentic AI systems, real-time data pipelines, and BI solutions across fintech, media, telecom, healthcare, real estate, and industrial engineering.
My work spans the full stack — from feature engineering and data mining through to deep learning with PyTorch, LLM orchestration with Claude, OpenAI, and LangGraph, causal ML, and live production deployment. I don't just build models — I build systems that keep working.
I've shipped real-time Kafka pipelines, multi-agent systems with Neo4j and Qdrant, and live products with paying users — all as an independent consultant working with founders and leadership teams worldwide.
Available as a full-time team member, part-time specialist, or contract partner — remotely, anywhere in the world.
💼 Let's Work TogetherNot a one-trick pony. Real systems built across multiple industries and problem categories.
I adapt to what your business needs — dedicated team member, specialist for specific hours, or project-based partner.
Join your team full-time — owning data strategy, infrastructure, and delivery end-to-end.
Flexible hours — ideal for companies needing senior data science expertise without a full-time hire.
Scoped engagements with clear deliverables — from single dashboards to full AI platforms.
Production-grade platforms with measurable impact — not toy models.
What it does: Predicts equipment failure before it happens and automatically generates a maintenance work order — no human in the loop.
Autonomous predictive maintenance platform using dual-head PyTorch Bi-LSTM + Attention on NASA turbofan data, routed through a 3-agent LangGraph loop with hybrid RAG (Neo4j + Qdrant).
What it does: Tells you who actually responds to a marketing campaign — not just who churns — so budget is never wasted on the wrong customers.
Causal ML platform on Criteo benchmark (2.5M rows) with S/T/X-Learner meta-algorithms, IPTW, and greedy budget optimiser. 3-container Docker deployment.
What it does: Reads live public text and turns it into structured sentiment data in real time — zero per-request API costs using a local LLM.
Real-time NLP pipeline: Hacker News API → Kafka → Ollama (Llama 3.2) → PostgreSQL → Streamlit. Resolved consumer eviction loops from 30–40s LLM inference latency.
What it does: A live, monetised trading alert product with real paying subscribers — not a demo.
Real-time WebSocket pipeline detecting price signals across multiple instruments, deployed on Linux VPS with Paystack subscription monetisation.
More Work
Every engagement follows the same rigorous process — so you always know what's happening and what comes next.
The full story behind the numbers.
Industrial equipment fails unexpectedly because traditional monitoring only flags problems after they've already happened. There's no system that connects fault detection to the fix automatically.
Built an autonomous platform ingesting raw sensor telemetry, predicting cycles remaining with a dual-head PyTorch Bi-LSTM + Attention network, then routing alerts through a 3-agent LangGraph loop. Agents query Neo4j + Qdrant hybrid RAG to retrieve maintenance procedures and check a SQLite parts inventory — producing a structured work order with zero human intervention.
Full pipeline from raw sensor input to structured work order in under 2 seconds. PSI drift engine correctly distinguished stable from drifted distributions across all 24 sensors.
View on GitHub →Standard ML tells you who will churn. That's not enough. If you contact a Sleeping Dog — a customer who churns specifically because you contacted them — you made things worse. Budget was being wasted on the wrong customers.
Built a causal ML platform on the Criteo benchmark (2.5M rows) implementing S-Learner, T-Learner, and X-Learner meta-algorithms with IPTW to correct selection bias. Evaluated using Qini coefficient and AUUC — not standard AUC. Deployed as a 3-container Docker system with a CEO budget simulator and MLflow experiment registry.
A system that separates Persuadables (target these), Sleeping Dogs (never contact), Sure Things (save budget), and Lost Causes (skip entirely). Budget allocated exclusively to customers who actually respond to intervention.
View on GitHub →Every project has a public GitHub repo — full code, READMEs, architecture diagrams, and results. No black boxes.
Tools I use in production — not just on tutorials.
Full-time, part-time, or contract — I'd love to hear about your project or role.
Download My CV