Open to remote roles globally — UTC+1
~/victor $
I build multi-agent pipelines, hybrid RAG architectures, and autonomous agentic workflows with evaluation infrastructure baked in from day one. Every system I ship runs in Docker, has documented tradeoffs, and answers the hard questions before they're asked.
Background
My foundation is Mechanical Engineering — failure analysis, constraint optimization, systems that fail in predictable and recoverable ways. When I moved into AI, I brought that same lens. I don't build demos that impress in a notebook and break in production. I build systems with documented tradeoffs, observable behavior, and clear failure modes.
My four deployed systems are built around the questions senior engineers actually ask: Why multi-agent instead of a single chain? How do you handle context accumulation across agent handoffs? What's your retrieval strategy when domain-specific terminology breaks semantic search? I have specific, defensible answers backed by real implementations.
I'm available for remote roles as an Applied LLM Engineer or Agentic Systems Engineer — with teams that are serious about production quality and want someone who can own an AI system end-to-end, from architecture decisions through deployment and continuous evaluation.
Selected work
Not prototypes. Containerized, evaluated AI systems — each built to represent a distinct engineering pattern and answer hard architectural questions about the decisions behind it.
Multi-agent fault diagnosis · LangGraph · 9 tools
A production-grade diagnostic platform built in four phases — from hybrid RAG retrieval over 27 industrial documents to full multi-agent orchestration with a Supervisor routing queries to four specialist agents. Features live MQTT telemetry, Gemini Vision for equipment image analysis, MCP server/client integration, and a custom 30-case evaluation suite. Every architectural decision — why multi-agent over a single chain, why hybrid retrieval over pure semantic search, why ChromaDB over a hosted vector DB — is documented and defensible.
LLM evaluation & regression platform · CI/CD integrated
Evaluation infrastructure for the Industrial AI Copilot — and the pattern that separates production AI engineering from demo-quality builds. An LLM-as-judge scorer evaluates RAG faithfulness, single-agent tool selection, and multi-agent synthesis quality across three distinct rubrics. A 5% regression in any mode triggers an automatic CI block before deployment. The gap between a one-time benchmark and continuous quality monitoring is exactly what AgentEval closes.
AI contract intelligence · CrewAI · 4-agent sequential pipeline
AI-powered contract analysis grounded in Common Law and UNIDROIT standards — demonstrating that the same agentic RAG architecture transfers to high-stakes legal domains. A four-agent CrewAI pipeline (Senior Analyst → Legal Researcher → Risk Assessor → Plain Language Specialist) produces structured risk reports with per-clause severity ratings and cited legal standards. Fourteen risk patterns detected automatically before any LLM call. Applicable across 80+ Common Law jurisdictions including UK, USA, Nigeria, and India.
Autonomous intelligence pipeline · LangGraph · schedule-driven
A fully autonomous agent pipeline that monitors industry news, scores articles for relevance, and synthesises structured intelligence briefings on a timer — no human trigger after deployment. Built on APScheduler + LangGraph with SHA-256 deduplication, four-category signal classification (Regulatory, Product, Market, Hiring), and a priority alert layer for breaking regulatory signals. Deployed for Nigerian fintech intelligence but domain-agnostic by design — point it at any industry by updating a single config file.
What I bring
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.
Most AI projects fail because the engineering solution gets built before the actual constraint is understood. I start there — cost, latency, data quality, retrieval reliability — and work backward to the architecture. That usually saves weeks of rebuilding.
Why multi-agent instead of a single chain? Why hybrid retrieval? Why this chunking strategy? I have specific, defensible answers — not intuition. That's what you need when a skeptical technical leader asks hard questions in a design review.
Everything I build is Dockerized, version-controlled, and connected to an evaluation pipeline. Not because I'm tidy — because I've seen too many AI projects die in Jupyter notebooks. The system should run the same on your infrastructure as mine, and you should know immediately when quality degrades.
Get in touch
Open to full-time remote roles, contract engagements, and consulting. If your team needs someone who can own an AI system from architecture to deployment — and walk through every decision clearly — reach out. I respond within 24 hours.