About
I help teams make confident AI infrastructure decisions by starting with the business case, user workflow, software architecture, and performance requirements before recommending hardware.
My work sits at the intersection of AI software, infrastructure, enterprise architecture, and practical delivery: private LLM deployment, CUDA and ROCm compatibility planning, GPU infrastructure planning, compute cluster setup, multimodal AI systems, agentic platforms, retrieval systems, and inference performance engineering.
What I Help Organizations Do
- Design enterprise AI platforms that are maintainable, scalable, and secure.
- Translate business use cases into software, infrastructure, and deployment requirements.
- Evaluate GPU, server, storage, networking, CUDA, ROCm, and software-stack decisions before large investments.
- Build and run PoC workloads on target hardware to validate real performance, compatibility, and deployment fit.
- Deploy private LLM and multimodal AI systems that keep sensitive organizational data under control.
- Improve inference throughput, latency, hardware utilization, concurrency, and user experience.
- Architect agentic AI platforms with governance, observability, explainability, and enterprise integration.
- Set up compute clusters when organizations need multiple machines working together.
- Build knowledge systems that connect search, retrieval, semantic understanding, and graph-based reasoning.
Business Case Before Hardware
Many vendors will sell a machine that can run selected LLMs. Enterprise AI planning is not that simple.
The right infrastructure depends on the actual business case: whether the system is a chatbot, an agentic workflow, a voice-to-text pipeline, a text-to-voice experience, a multimodal assistant, a VR integration, a retrieval system, a fine-tuning workflow, or a high-concurrency internal platform.
Before hardware decisions, I help teams clarify expected users, concurrency, latency targets, model architecture, inference patterns, training or fine-tuning needs, framework choices, data sensitivity, deployment model, and integration requirements.
That includes understanding CUDA, ROCm, driver support, inference frameworks, model-serving requirements, and the software dependencies an organization may not realize will shape the hardware decision. Software architecture comes first. Hardware follows from the workload.
Enterprise AI Architecture
My work focuses on the engineering layer between AI ambition and production reality.
Current work includes enterprise architecture for agentic AI platforms: multi-agent systems, tool orchestration, governance, explainability, observability, deployment architecture, and scalable AI platform design.
The goal is practical architecture: systems that teams can operate, executives can trust, and organizations can extend without creating expensive technical debt.
Production AI Systems
Production AI requires more than a model demo. It needs a deployment path, infrastructure plan, performance targets, monitoring, security boundaries, and a clear operating model.
I help organizations move AI systems toward production readiness across private LLM deployment, production inference, asynchronous pipelines, scalable APIs, multimodal workflows, and enterprise integration.
Technologies may include modern inference servers, containerized services, API layers, retrieval systems, orchestration frameworks, and GPU-backed infrastructure, but the focus is always the operational outcome.
AI Infrastructure Advisory & Procurement Support
AI infrastructure decisions can become expensive quickly. I help organizations evaluate options before committing to hardware, cloud spend, or vendor proposals.
- Proposal evaluation and pricing validation.
- GPU, CPU, RAM, storage, networking, and cluster design assessment.
- AI workstation and server planning.
- Workload sizing for inference, training, fine-tuning, retrieval, agentic workflows, speech systems, and multimodal workloads.
- CUDA, ROCm, driver, framework, and model-serving compatibility review.
- Infrastructure benchmarking and performance assumptions.
- PoC workload design and testing on target hardware before procurement or rollout.
- Software stack and framework evaluation.
- Architecture recommendations for cloud, on-premises, or hybrid deployments.
- Multi-machine compute cluster setup and deployment planning.
This work is vendor-neutral. I have evaluated solutions involving Dell, Lenovo, HP, NVIDIA ecosystem technologies, and related enterprise AI infrastructure options without implying endorsement of any vendor.
AI Performance Engineering
Many organizations do not need new hardware first. They need better use of the infrastructure they already have.
I help improve AI serving performance by reducing latency, increasing throughput, improving concurrency, and maximizing GPU utilization. Practical work may involve continuous batching, asynchronous inference, KV cache behavior, high-concurrency serving, distributed inference patterns, cluster-aware deployment, and inference pipeline design.
In appropriate workloads, I have helped improve hardware utilization by 10x to 250x through better parallel inference architecture, batching, and deployment design.
Government & Enterprise Experience
Government and enterprise environments require clear communication, explainability, reliability, and respect for operational constraints.
They also follow specific budget cycles, procurement processes, procurement vehicles, Government of Canada pricing expectations, timelines, and bidding requirements. Technical recommendations need to fit that reality.
I have advised and built AI, analytics, and infrastructure solutions in large Canadian government environments, including work with the Canada Revenue Agency and other federal organizations.
I also contributed to a first-place Public Service Data Challenge team that designed an early federal generative AI chatbot, helping demonstrate how hybrid search and GenAI could improve access to public-service information.
This background helps me bridge the gap between hands-on engineering detail and the decision-making needs of CIOs, CTOs, directors, procurement teams, technical founders, and delivery leaders.
Typical Engagements
AI Infrastructure Assessment
Assess whether current or proposed infrastructure can support target AI workloads, including user demand, concurrency, latency, inference demand, storage, networking, security, and operational readiness.
GPU Procurement Review
Review hardware proposals, pricing assumptions, sizing logic, CUDA/ROCm compatibility, software requirements, cluster requirements, and deployment risks before major infrastructure purchases.
AI Architecture Review
Evaluate an existing or planned AI platform for scalability, maintainability, security, deployment readiness, and integration risk.
Enterprise Agentic AI Roadmap
Design a practical roadmap for agentic systems, including use cases, architecture, governance, observability, tool orchestration, and rollout sequence.
Private LLM Deployment
Deploy secure AI systems that keep sensitive organizational data under organizational control while supporting usable internal workflows.
Production Inference Optimization
Increase throughput, reduce latency, improve concurrency, and maximize hardware utilization before investing in additional infrastructure.
Enterprise RAG Strategy
Design retrieval systems that connect organizational knowledge to AI assistants and search tools with better accuracy, traceability, and governance.
Knowledge Graph Design
Represent relationships across documents, systems, entities, and workflows so teams can reason across complex enterprise information.
AI Benchmarking
Compare models, hardware, software stacks, and deployment options against practical workloads rather than relying on vendor claims alone.
PoC Workload Testing
Build and run proof-of-concept AI workloads on target hardware so teams can validate throughput, latency, compatibility, utilization, and operational fit before procurement or production rollout.
Why Work With Me
- Enterprise AI architecture experience.
- Production deployment and integration experience.
- Vendor-neutral AI infrastructure advice.
- CUDA, ROCm, framework, and hardware-stack compatibility judgment.
- Government and enterprise delivery context, including CRA experience and work with other federal organizations.
- First-place Public Service Data Challenge GenAI chatbot experience.
- AI procurement support.
- PoC workload testing on target hardware.
- Public LLM hardware benchmarking and visualization.
- Performance engineering and inference optimization.
- Open-source engineering.
- Executive communication.
- Hands-on implementation.
Selected Case Studies
Confidential Enterprise Multimodal AI Architecture
Challenge: Enterprise AI workflows increasingly combine speech recognition, language models, speech synthesis, immersive interfaces, and application integration. These systems must feel responsive while still meeting infrastructure and deployment constraints.
Approach: Designed and optimized multimodal AI architecture that connected speech, language, synthesis, interface, and inference components into a unified workflow.
Outcome: Improved the path from prototype components to an integrated enterprise AI workflow with clearer architecture and performance considerations.
Public Service Data Challenge GenAI Chatbot
Challenge: Public-service information can be difficult to navigate when users ask long, specific, real-world questions.
Approach: Contributed to a first-place Public Service Data Challenge team that designed an early federal generative AI chatbot using hybrid search and retrieval-oriented design.
Outcome: Helped demonstrate how GenAI could improve discovery and access to government information while keeping the system grounded in relevant source material.
Private LLM for Program Evaluation
Challenge: Evaluation teams needed a way to explore private AI for sensitive analytical work while understanding what production readiness would require.
Approach: Built a retrieval-augmented local LLM prototype and translated infrastructure, MLOps, containerization, and CI/CD considerations for technical and non-technical leaders.
Outcome: Helped stakeholders understand how private AI systems could support evaluation workflows and what would be required to operate them responsibly.
Enterprise Search and Knowledge Systems
Challenge: Stakeholders needed faster ways to navigate complex organizational information and high-dimensional enterprise data.
Approach: Developed search, summarization, and mapping tools that made large internal data sources easier to inspect, query, and explain.
Outcome: Improved discovery workflows and supported more efficient analysis of enterprise information assets.
AI-Assisted Contract Review
Challenge: Legal and compliance documents can be difficult for non-specialists to interpret consistently.
Approach: Developed an online contract review tool that parses PDF rental agreements and uses LLM-supported analysis to flag potential issues for review.
Outcome: Created a practical compliance checker with an independently managed production workflow.
Selected Technical Work
These projects are included as evidence of hands-on implementation, not as a catalogue of repositories.
AI-Assisted Contract Review Platform (Closed Source)
What it demonstrates: End-to-end AI product delivery, from document ingestion and backend logic to web application deployment.
Engineering work: Built an online review tool for rental agreements using PDF parsing, LLM-supported analysis, a Python backend, a Next.js frontend, Linux server deployment, and CI/CD.
Why it matters: Shows the ability to turn an AI workflow into a usable production application with real users, not just a notebook or prototype. The codebase is closed source.
LLM Hardware Benchmark Visualization
What it demonstrates: Benchmark-driven hardware evaluation for LLM inference, including performance, memory bandwidth, and cost/performance tradeoffs.
Engineering work: Visualized LMSYS benchmark data comparing RTX Pro 6000 and DGX Spark results across model and batch-size scenarios.
Why it matters: Shows how I evaluate AI hardware using experimental results and workload behavior instead of relying only on vendor positioning or headline specifications.
Early Local RAG Prototype
What it demonstrates: Early hands-on work with local retrieval-augmented generation for sensitive program-evaluation workflows.
Engineering work: Built local RAG patterns two to three years ago, when the surrounding frameworks were far less mature and more of the system design had to be assembled manually.
Why it matters: Shows long-running experience with private AI patterns, retrieval design, and workflow fit before local LLM tooling became common.
LLM on Google Colab Tutorial
What it demonstrates: Practical teaching material for running LLM experiments in constrained cloud notebook environments.
Engineering work: Wrote a tutorial for setting up LLM workflows in Google Colab before local and hosted LLM tooling became as accessible as it is now.
Why it matters: Shows the ability to explain AI infrastructure concepts clearly and help others reproduce technical workflows.
Recursive SQL for High-Dimensional Tables
What it demonstrates: A tutorial-style exploration of recursive SQL for summarizing high-dimensional database tables.
Engineering work: Developed examples for inspecting complex table structures and producing summaries directly on database servers.
Why it matters: Shows depth in data infrastructure and database reasoning, which matters when AI systems depend on messy enterprise data.
Blog
Technical writing and notes are available at Journal Learn.
Future topics should focus on practical enterprise AI questions such as private LLM deployment, GPU sizing, AI procurement, inference optimization, enterprise RAG, and agentic AI architecture.
Contact
Whether you're evaluating AI infrastructure, deploying private LLMs, building enterprise agentic systems, or looking to optimize existing AI platforms, I'd be happy to discuss your project.
- Name: Henry Luan
- Email: info@sonarwave.org
- Website: www.bigkpi.com
- GitHub: casualcomputer
- Blog: Journal Learn