AI Engineer & Systems Architect
I build production AI systems that connect models, agents, organizational knowledge, infrastructure, and real-world workflows.
Enterprise AI architecture connecting private models, organizational knowledge, agents, and internal systems — ingestion, indexing, retrieval, and LLM reasoning with agents that search documents, query databases, and produce reports.
An AI creative director for small businesses and creators. From one plain-language brief, the agent plans the content, shows the credit price, waits for approval, produces every piece — images, video scenes, voiceover, music, captions — assembles the final video, and publishes it to Facebook, Instagram or TikTok once a human approves. Multi-model, budget-aware, and strong in Burmese and Thai.
A driving policy trained with reinforcement learning in simulation and transferred to a real computer-controlled electric golf cart — a hands-on study in sim-to-real transfer.
As Senior AI Developer, Min works across the complete lifecycle of an AI system: identifying business problems with leadership, selecting architectures, designing databases and backend services, building user-facing applications, integrating models and agents, deploying GPU infrastructure, and controlling cost — building an interconnected internal AI environment rather than isolated tools.
An internal intelligence system allowing department heads to query organizational knowledge and operational data, grounded in internal documents and structured business data.
Converts fragmented notes, documents, and spreadsheets into structured investor intelligence: interactions, milestones, status, and next actions.
Turns venture-capital information into structured intelligence with company tracking, risk indicators, and AI-generated summaries.
Makes employee career development structured and data-driven by connecting profiles, skills, career interests, roles, learning opportunities, and skill gaps.
Built from scratch: employee profiles, leave and calendar workflows, and process automation, with a long-term direction of AI-assisted skills intelligence.
A core engineering belief shows up here: not every problem requires AI, and an AI system still needs strong conventional software underneath it.
Primary technical owner of AI, reporting to the CEO. Production AI across internal organizational workflows, private LLM infrastructure (DGX-class GPUs, vLLM, benchmarked on tokens/sec, time-to-first-token, and concurrency), knowledge systems, agents, and full-stack applications — with cloud cost treated as an engineering problem, not an accounting one.
Depth-camera trench measurement system for Thailand’s Metropolitan Waterworks Authority — point clouds, measurement algorithms, Jetson field deployment.
Taught machine learning, Python, web development, IoT, robotics, embedded systems, and computer vision through hands-on projects.
Building AI-native software where agents are part of the product architecture, not an add-on.
Founded in 2026 and incorporated in Delaware, SoeMindAI is a product company. Its product is SoeMind: an AI creative director that plans, prices, produces and publishes marketing content for small businesses, creators and agencies. It grew out of Min’s open-source SoeMind Forge toolkit.
You direct. The agent plans, prices and produces.
Most generative AI tools are single-step: one prompt, one output. Real marketing production is a pipeline. In SoeMind you describe a goal — “a 15 second Reel for my bakery’s new croissant” — and the agent asks the few questions that matter, writes the script and scene plan, shows a production plan with the credit price of every step, and generates nothing until you approve. Then it produces the images, video scenes, voiceover, music and captions, assembles the video, checks it against the brief, and can draft and schedule the post.
Every paid step shows its price before it runs and waits for approval, or runs inside a budget you set. Redo one scene without paying for the whole video. A post goes live only after a human approves it. Your brand — colors, tone, logo, characters, products, voice — lives in a Project and is reused automatically.
Chat with the AI director. It asks only what it needs, writes the idea and scene plan, prices every step in credits, and produces after one approval. Auto-run mode works inside a budget you set.
Hands-on tools for one asset at a time — image, video, talking avatar, lip-sync, voiceover, voice clone, music, sound effects, transcription, captions, re-dubbing — with the price on the button.
About a hundred proven formats (what-if ads, testimonials, founder stories, explainers): one short form, one price, one approval.
Publish and schedule to Facebook Pages, Instagram and TikTok, read insights on what works by hook, topic, format and timing, and plan the next campaign from the evidence. Nothing posts without a human approving it.
SoeMind is also an MCP server, so Claude, ChatGPT or any MCP client can plan and produce content in a user’s account — every paid action still confirmed by the user.
Model-agnostic by design: the agent picks per scene across Veo, Seedance, Kling, Gemini Omni, Nano Banana, GPT Image and more, on quality, cost, speed and task. Built for English, Burmese and Thai first — localization covers script, voice, captions, presenters and visuals, not just translation — with the interface available in 19 languages. Credits are priced before anything runs; failed generations are refunded automatically.
Trained an autonomous driving policy with reinforcement learning (PPO) and UNet/ResNet semantic segmentation in the CARLA simulator, then transferred it to a computer-controlled electric golf cart with real steering, throttle, and braking. A hands-on lesson in sim-to-real transfer: what works in simulation often breaks against real noise, lighting, and mechanical limits.
A depth-camera system for water-pipe trench quality control: point-cloud preprocessing, coordinate transformation, trench profiling, and measurement algorithms, achieving errors below roughly five percent and deployed on an NVIDIA Jetson Xavier NX for field use.
This background feeds a continuing interest in Physical AI: connecting intelligence to sensors, actuators, robots, and vehicles — Jetson, Raspberry Pi, STM32, ESP32, Arduino.
Project-first teaching focused on taking students from intuition to implementation and experimentation — an image-classification project before the equations. That mentoring has reached publication level: Min guided a 16-year-old Thai student through a research project that became a peer-reviewed IEEE paper, “Reinforcement Learning application on Gamification”, presented at the 2023 8th International STEM Education Conference (iSTEM-Ed).
Python, TypeScript, JavaScript, C++, Go
LLMs, AI agents, RAG, PyTorch, TensorFlow, deep learning, reinforcement learning, generative AI, multimodal AI
GPT, Claude, Gemini, Llama, Qwen, Hermes
YOLO, SAM, ResNet, UNet, OpenCV, segmentation, object detection, point clouds
Veo, Seedance, Kling, Gemini Omni, Nano Banana, GPT Image, FLUX, image-to-video, lip-sync, text-to-speech, voice cloning, AI music
FastAPI, REST APIs, React, Next.js, Vite, Tailwind CSS
PostgreSQL, MySQL, Redis, AWS, Docker, NGINX, RunPod, GPU deployment, private AI infrastructure
NVIDIA DGX, NVIDIA Jetson, Raspberry Pi, Arduino, STM32, ESP32
An impressive demo is not enough. The question is whether the system saves time, improves decisions, removes repetitive work, or creates a capability that did not exist before.
A language model alone is not an agent. The model provides reasoning and planning; databases, APIs, search, document systems, and generation models provide the ability to act.
Autonomy should not automatically mean removing humans. For high-value decisions and expensive generation tasks, the best pattern is: human objective, AI plan, human review, AI execution, human feedback.
Models change quickly. Product logic, agent logic, and model providers should stay separated so models can be swapped as the technology improves.
Production AI needs reliable APIs, security, observability, latency control, and fallback systems. A powerful model on poor infrastructure is still a poor product.
GPUs, model APIs, and generation pipelines get expensive fast. Cost should shape model selection, caching, retry strategy, and architecture from day one, not appear later as an accounting surprise.
Alongside these sits Responsible AI: human oversight, privacy and organizational data protection, reliability, transparency, and hallucination risk — especially in enterprise settings that handle confidential information.
The future is not agents instead of software engineering. It is software engineering + agents + intelligent interfaces.
Interested in enterprise AI, agent systems, multimodal products, or AI infrastructure?
Book a conversation