Software · AI systems · Products
Software that makes complex systems simpler.
I'm Phu, an independent developer in Bangkok. I build software that connects artificial intelligence, business operations, and practical engineering.
I build complete systems, not just screens: the data model, the API, the AI that reads it, the interface people use, and the operations around it. Most of what I make starts with a real business and a real problem.
Selected work
Four systems, in depth.
Each one is told in three parts: the idea, what is built, and what is still open.
01 · Full-stack platform
Gumpun
One platform for a Muay Thai school's customers, staff, and visa cases.
In progress: Internal pilotA customer asks
Training and visa questions are answered from the school's own knowledge, not from a model's guess.
The agent works through tools
Lookups go through read tools. Changes go through three controlled actions. The agent can only do what it has been given.
Staff stay in charge
Anything that needs judgment is handed to a person, and lands in the same workspace as the student's record and visa case.
02 · Enterprise AI
ERP AI
Ask a business question in plain language. Get an answer from the ERP.
In progress: Phase 1 builtThe idea
Someone asks what revenue was last month, or how two periods compare, without opening a single ERP table.
What is built
A synthetic PostgreSQL ERP and a read-only API with sales-summary and period-comparison endpoints, under automated tests.
What is still open
The model. A local 14B model did not meet my acceptance criteria, so the AI layer is the next phase, not a finished one.
03 · Operational intelligence
Business World
ERP records, shown as an interactive world instead of a dashboard.
Implemented: Experimental buildOperations as a place
Business entities become things you can see and inspect, with their relationships intact.
Follow a lot forward
In the rice-mill profile, one paddy lot becomes a milling batch and then several outputs, each with its yield.
Trace it back
Pick any output and walk it back to the lot it came from. That is the basis for recall analysis.
04 · Developer tooling
Fox's Shelter
A command center for AI agents, projects, and the machines they run on.
Implemented: Working prototypeEvery agent, one view
Activity state, session, task status, and availability for each development agent.
The state of the work
Repository status, working-tree changes, test results, and build status beside the agents producing them.
The machine underneath
CPU, memory, disk, and load over time, plus an approval inbox for the actions that should wait for a yes.
How I build
The whole system, end to end.
I design the system, write the specification, and direct AI coding agents through implementation. I dispatch their work deliberately and keep them supervised. Architecture, test requirements, and release decisions stay with me. I say so plainly because that is how I actually work, and a portfolio should be honest about who did what.
01
Data
PostgreSQL schemas, controlled views, and read-only access where it matters.
02
API
Validated, tested endpoints that expose only what a consumer should reach.
03
AI
Local and hosted models, tool calling, and agents with narrow permissions.
04
Interface
React interfaces designed around the people doing the work.
05
Operations
Deployment, monitoring, approvals, and a human review loop.
Skills
Sorted by how well I know them.
A flat list of technologies says very little. These are grouped by how much I have actually done with each.
Hands-on
Used in work shown on this site.
- Python and FastAPI
- TypeScript, React, and Next.js
- PostgreSQL and SQL
- REST API design
- Git and GitHub
- Docker
- Local LLM inference (LM Studio, llama.cpp)
- Directing AI coding agents (Codex CLI)
Working knowledge
Comfortable, still deepening.
- Node.js services
- Relational data modelling
- Automated, acceptance, and end-to-end testing
- CI workflows
- Tool calling and agent design
- Requirements analysis and workflow design
- Linux, WSL2, macOS, and Windows environments
- SSH and Tailscale networking
- ERP and inventory concepts
Exploring
Reading, testing, not yet shipped.
- Quantization and GGUF model tuning
- Multi-machine inference
- Agent observability
- Cloud versus local inference, and what each costs
- Multi-tenant data isolation
- Role-based access for AI systems
- Odoo integration
In development
Not built yet, and labelled that way.
Packaging wholesale management system
A proposed business management system for a packaging wholesale operation: inventory, accounting workflows, administration, and internal records in one web interface. Requirements and development planning so far; nothing built yet.
Not every business needs an AI agent. Some need reliable inventory tracking, better accounting workflows, and software their staff will actually use.
Off the keyboard
Precision, in other forms.
Practical shooting
A background in competitive practical shooting, which shaped how I think about precision, discipline, and performance under pressure.
Brazilian Jiu-Jitsu
Training regularly. Currently a four-stripe blue belt.
Hardware and networking
Building machines, wiring up a home lab, and running models locally. Some gaming too.
Coffee
Taken seriously, usually while something compiles.
Notes
Writing it down.
Contact
Have a system worth building?
I'm interested in real business problems, especially where AI has to work with existing data and existing people.
requiemsen@dhanyathep.site