Available for selected AI engineering projects — UTC+1
AI systems that
do real work.
RAG · AI Agents · Workflow Automation · Evaluation
I help teams automate knowledge-heavy and operational workflows — research, documentation, diagnostics, reporting — with AI systems built for real-world operation.
Evaluation, observability, failure handling, and deployment are considered from the beginning — not added after the demo.
30-case eval suite
Problems I solve
What can AI take off your team's plate?
I build systems around workflows where teams spend too much time searching, researching, analyzing, classifying, reporting, or moving information between tools.
Knowledge & RAG Systems
Turn internal documents, technical knowledge, SOPs, and research material into AI systems your team can actually use.
Intelligent Workflow Automation
Automate repetitive research, classification, monitoring, reporting, and decision-support workflows.
AI Agents & Decision Workflows
Connect AI models to tools, data, and business processes to execute structured multi-step work.
AI Evaluation & Reliability
Measure AI quality continuously and catch regressions before they reach users.
Before we build anything
AI Workflow Assessment
I'll help determine whether AI is actually a good fit for your workflow — and what I would build if it is. Not every problem needs an agent; some need a script, and some don't need AI at all.
Why I build this way
I approach AI systems as engineering systems — not experiments.
My foundation in Mechanical Engineering trained me to think in terms of constraints, failure modes, tradeoffs, and system behavior. I brought that mindset into AI engineering: I design systems with real-world operation in mind — measurable evaluation, failure handling, observability, and reproducible deployment, considered from the start.
Each system in my portfolio started with a concrete workflow problem, not a technology I wanted to try. That means I can walk through why a given architecture was chosen, where it would break, and what changes under a different constraint — the kind of scrutiny a serious technical buyer or hiring manager should apply.
I'm available for AI engineering projects, contract work, and remote roles — with teams and founders who want someone who can own a system end-to-end, from diagnosing the workflow through architecture, deployment, and evaluation.
University of Uyo
Systems I've built
Proof through working systems.
Each system started with a concrete problem, an architectural constraint, and a measurable definition of success.
Industrial AI Copilot
AI-assisted industrial knowledge and fault diagnosis
Industrial troubleshooting can require searching technical documentation, interpreting equipment information, and combining multiple sources of evidence.
Built a multi-agent diagnostic system combining hybrid retrieval, specialist agents, live telemetry, image analysis, and tool integration.
Domain-specific terminology made pure semantic retrieval insufficient, leading to a hybrid dense + BM25 retrieval strategy with reranking.
Migrated from Hugging Face Spaces to AWS: containerized on ECS Fargate behind an Application Load Balancer, with S3-backed vector store persistence and a GitHub Actions CI/CD pipeline authenticated via OIDC — no long-lived AWS credentials stored in CI.
30-case eval suite · 90% answer accuracy · 100% retrieval precision · 3.09s average latency
Busiv
Autonomous business intelligence workflow
Monitoring industry news and producing structured intelligence can consume hours of manual research every week.
Collect → filter → classify → deduplicate → synthesize → deliver. An autonomous pipeline that produces scheduled intelligence briefings — no human trigger required after deployment.
Raw news volume is mostly noise. A relevance-scoring and four-category signal layer (Regulatory, Product, Market, Hiring) filters what reaches synthesis, with a priority path for time-sensitive regulatory signals.
AgentEval
AI quality and regression infrastructure
AI systems can silently degrade as prompts, models, retrieval strategies, or orchestration logic change.
Built an evaluation framework for RAG, single-agent, and multi-agent workflows with LLM-as-judge scoring and CI/CD regression gates.
Provides a repeatable quality signal before changes are deployed.
LexAI
AI-assisted contract intelligence prototype
Contract review involves extracting clauses, identifying risk patterns, and presenting complex language in a structured form.
Built a multi-agent document analysis pipeline (Senior Analyst → Legal Researcher → Risk Assessor → Plain Language Specialist) with structured risk extraction and domain-grounded retrieval against Common Law and UNIDROIT reference material.
Research prototype — not legal advice.
Not ready to talk yet? Read the flagship case study and see the architecture behind the numbers.
See the Case Study →Flagship case study
From fragmented documentation to an AI-assisted diagnostic workflow.
Industrial troubleshooting means searching manuals, reading equipment telemetry, and interpreting visual evidence — usually across three separate tools. The Industrial AI Copilot does this in one conversation.
Domain-specific terminology made pure semantic retrieval insufficient. Combining dense vector search with BM25 keyword matching, then reranking, closed the gap between what the model retrieved and what a technician actually needed.
Vision analysis depends on image quality from the field; telemetry integration currently assumes MQTT-compatible sensors. Both are documented, not hidden, in the source repository.
The same pattern — hybrid retrieval + live data + specialist agents + evaluation — applies to any operational workflow combining documentation, sensor data, and visual evidence.
Originally deployed on Hugging Face Spaces, later migrated to AWS — ECS Fargate, ECR, an Application Load Balancer, S3-backed persistence, and a GitHub Actions pipeline authenticated via OIDC rather than stored credentials. The migration surfaced a real production lesson: an ALB health check race condition, fixed by separating liveness from readiness so the service only reports healthy once retrieval and agents are actually loaded.
Engineering
How I build AI systems.
A mechanical engineering foundation means I think about systems the way systems actually behave — with edge cases, failure modes, and constraints that matter in the real world.
Architecture follows the constraint, not the other way around.
Cost, latency, data quality, reliability, and workflow complexity determine architecture. I start there and work backward, which usually saves weeks of rebuilding.
Use a single chain when a single chain is enough.
Introduce agents only when they provide a real capability advantage. Reaching for multi-agent orchestration by default is a tell of inexperience, not sophistication.
A system that hasn't been measured hasn't been proven to work.
I define measurable quality criteria and regression tests before assuming the system works — not after users start noticing problems.
Deployment is part of the system, not cleanup work.
Containerization, observability, failure handling, versioning, and reproducible deployment are considered from the start, not bolted on before a demo. The flagship system runs on AWS ECS Fargate with an automated, OIDC-authenticated CI/CD pipeline — not just a hosted demo link.
Who I help
Where I can create leverage.
Founders building AI products
You have a validated problem and need someone who can turn the workflow into a reliable, production-grade AI system.
Teams drowning in information
Your people spend hours searching documents, researching markets, preparing reports, or processing information manually.
Companies with an AI prototype
You've built something that works in a demo but need better retrieval, evaluation, orchestration, or deployment to trust it in production.
Get in touch
Have an AI workflow worth automating?
Tell me what your team currently does manually, where the bottleneck is, and what you've already tried. I'll help determine whether AI is actually a good fit — and what I would build if it is.