Min Liu

Open To

AI systems and agentic application development, C++ software engineering and high-performance ML systems, developer tooling, and backend/data infrastructure roles with teams in British Columbia, across Canada, or remote-friendly North American organizations.

Technical Strengths

• Modern agentic systems, end to end: LLM fine-tuning, agent loops, orchestrators, MCP tool protocols, skills with progressive disclosure, human approval gates, session observability, and evaluation loops — the full stack that makes an agent production-ready instead of just impressive.
• C++ high-performance ML systems: recommendation and ranking engines, feature pipelines, continuous training and model serving, and experiment design — built over more than six years of production machine learning engine work.
• Applied AI products: RAG workflows, voice/transcription pipelines, structured outputs, dashboards, and production LLM pipelines.
• C++ developer tools: Qt/QML desktop tooling, Clang/LLVM APIs, source-code analysis, lazy rendering, metadata extraction, and practical debugging workflows.
• Backend and data infrastructure: Python services, data pipelines, structured storage, queues, retry handling, observability, and production reliability.

Selected Work

Data infrastructure · LLM pipelines · product reliability

Actual Voice

Voice intelligence platform for teams that need to understand open-ended feedback without manually reading every transcript or exposing raw individual responses.

The user value is not transcription by itself. The useful outcome is turning messy spoken responses into patterns that leaders can act on: engagement signals, recurring themes, and dashboard-ready summaries that preserve the difference between individual raw data and aggregate insight.

This kind of product depends on both product judgment and infrastructure discipline. The pipeline has to handle uploads, transcription, LLM analysis, scoring, aggregation, retries, tenant boundaries, and dashboard contracts while keeping the experience simple for the people reviewing the results.

Built with Supabase Storage/Postgres/Edge Functions, worker queues, Whisper, LLM classification, dashboards, retry handling, observability, tenant-aware limits, and production pipeline hardening.

Agentic AI · RAG knowledge engine · equipment operations

OncoMate

AI workspace for oncology equipment operations with an agentic backend: a technician describes an equipment fault in plain language, and a multi-turn agent loop retrieves fault-code references, similar completed repairs, PM notes, and parts stock to turn the question into a reviewable work plan.

The assistant is useful because the product knows the equipment. A knowledge engine ingests PMI machine files (.xml/.wox) into structured records — fault events, flow states, board temperatures — and folds every completed work order, maintenance result, and service note back into the retrieval layer. When an issue comes in, the agent runs RAG search over that organizational history plus system knowledge, then works in a loop: match the asset, review prior fixes, check schedules and inventory, and propose grounded next checks.

I designed the system around the constraints of regulated operational work: authenticated users and tenant boundaries, retrieval grounded in source records rather than free generation, and evidence captured when work closes — so each finished repair makes the next answer better.

Built with Next.js App Router, Supabase Auth/Postgres/Edge Functions, OpenAI APIs, a multi-turn tool-calling agent loop, RAG over maintenance/repair history and fault-code knowledge, PMI file parsing with metric trend views, and Tailwind.

Developer tools · C++ · compiler infrastructure

ACAV

A C++/Qt desktop tool that helps engineers, researchers, and students inspect Clang AST structure without reading raw compiler dumps.

Compiler tooling is powerful, but the normal debugging experience is often hostile: huge AST dumps, hard-to-follow source locations, and little help connecting compiler facts back to the code a person is trying to reason about. ACAV turns that into an inspectable visual workflow.

The project reflects the kind of engineering I want to keep doing: C++ systems work, compiler APIs, structured data extraction, UI performance, and tools that make complex technical evidence easier to use.

Built with C++, Qt 6/QML, Clang/LLVM AST APIs, source-location indexing, lazy tree rendering, structured metadata extraction, diagnostics, and GitHub Pages documentation.

Developer tools · C++ · JSON parsing

jqcpp

A C++20 command-line JSON processor that parses JSON, evaluates jq-like filters, and prints readable structured output.

JSON is easy for software to exchange, but it is often awkward for a person to inspect quickly. jqcpp focuses on the core workflow developers use in the terminal: pull out a field, read one array item, slice a response, or ask for keys and lengths without writing a temporary script.

The project is intentionally compact, which makes the engineering visible. It separates tokenizing, parsing, expression AST construction, evaluation, pretty printing, command-line input, and test coverage instead of hiding the behavior inside one large string-processing path.

Built with C++20, CMake, a JSON tokenizer/parser, jq-like expression lexer/parser, AST evaluator, vector-backed ordered objects, pretty printing, Catch2 tests, and GitHub Actions CI.

Who I Am

My work stands on two pillars.

The first is modern agentic systems. I build production-ready agents end to end — LLM fine-tuning, agent loops, orchestration, MCP tool integration, skills, observability, and the evaluation harnesses that prove an agent actually works. The projects above are complete agentic products, not demos: Chamfer turns text or images into editable 3D models through a visible, self-checking design loop, and OncoMate runs a multi-turn agent loop over a RAG knowledge engine in a regulated operational domain.

The second is high-performance machine learning systems in C++. I spent more than six years building recommendation systems and production ML engines at scale: at Alibaba I worked on CTR/CVR ranking, causal marketing science, SEO automation, feature pipelines, and model-serving workflows behind real traffic. That background is why my agent work treats reliability, observability, and evaluation as first-class engineering rather than afterthoughts.

Today I am building portfolio-grade systems and products in British Columbia while focusing my job search on AI systems and agentic development, C++ software engineering, developer tooling, and data infrastructure roles.

Capabilities

• Production-ready agentic systems: LLM fine-tuning, agent loops, orchestration, MCP tool integration, skills, measured verification, observability, and evaluation.
• C++ high-performance ML systems: recommendation and ranking engines, feature pipelines, continuous training and model serving — plus Qt/QML desktop tooling and Clang/LLVM APIs.
• Backend and data infrastructure: Python services, data pipelines, structured storage, queues, retry handling, observability, and production reliability.

Stack

C++ · Python · Linux · LLM Agents · LLM Fine-tuning · MCP · Recommendation Systems · Model Serving · Qt/QML · Clang/LLVM · CMake · PostgreSQL · Docker · Data Pipelines · Worker Queues · Observability

Contact

For conversations about C++ systems, developer tools, backend infrastructure, or data infrastructure work, LinkedIn is the easiest place to reach me.

Best first contact: LinkedIn · Code and projects: GitHub