The era of merely consuming AI is over.
We design and provide it.
Calling an API is not the same as building an AI system
Anyone can call a model. Yejua builds and runs the layer above it — infrastructure, orchestration, reliability and cost.
Multi-provider LLM infrastructure
We built and operate our own gateway that governs OpenAI, Anthropic, Gemini and AWS Bedrock behind a single endpoint — issued API keys, usage ceilings, per-model cost accounting and automatic monthly settlement included.
Every LLM call across our products passes through it
Cross-model verification
Important judgements are never left to one model. Separate models analyse independently and a judging model reconciles them; elsewhere, agents arguing from conservative, aggressive and cost positions debate before a referee concludes.
Running in report generation and internal decision-making
Agent design with a human in the loop
High-stakes calls are not pushed through automatically. At the point a human decision is needed, a signed-token link goes out, and the paused task resumes from exactly where it stopped once the answer arrives.
Verified end to end: request → reply → task resumption
Handling each provider's sharp edges
The same prompt breaks differently on each model. Models that reject a temperature parameter; models whose reasoning tokens eat the output budget until the JSON is truncated. None of this is in the docs — it surfaces in production. Our adapter layer absorbs it.
Swapping models does not require touching call sites
Hallucination control
An AI writing commerce copy that invents a specification has produced false advertising. Anything absent from the brief's whitelist never reaches the copy, and Korean text is rendered in HTML rather than generated as an image.
A founding design rule of our detail-page pipeline
Image understanding and generation
Vision — OCR and image comprehension — across every provider, plus mask-based editing and image costing derived from resolution and quality. Foreign-language text inside product images is re-rendered in Korean with its position and styling preserved.
Includes generative image work built on Azure OpenAI
Built by people who have done the work
Built on two decades of engineering depth, and on the frustrations we met selling online ourselves — which is what the automation was written to solve.
Tools go wrong when you don't know the floor
Subtract fees and international shipping and the margin disappears. Detail pages have to be rebuilt for every language. Order status is never where the API docs say it is. None of that reaches your design unless you have sold something yourself.
We don't build what we wouldn't use
Every Yejua product runs on our own operations first. The commerce automation drives our own store; the AI Gateway carries every LLM call across our products. We verify before we sell.
We tell you who builds it
The largest unknown in software is who actually writes the code. At Yejua the founder designs and implements it.
Choi Hyungjun
Founder · CTO
- Exp.
- 20+ years, full-stack
- Role
- Architecture · build · operations
- contact@sales-agent.kr
Nothing is handed down a chain
When sales takes the brief and passes it to a developer, requirements get translated twice and arrive bent. Here, one person's judgement runs from architecture through implementation, deployment and incident response. Decisions are fast, and what ships actually runs.
Track record
- Enterprise systemsLed large-scale system design and build across multiple IT organisations, resolving technical debt and growing teams alongside it
- Enterprise cloudOperated VM-based services on Microsoft Azure, built and migrated Azure SQL databases, and shipped generative image features on Azure OpenAI
- Selling, first-handSold online directly, met the limits of sourcing, margin and inventory, and solved that stretch with automation
- Own SaaS in productionBuilt and operates an LLM gateway and a line of agent and commerce products on AWS
How we got here
- Foundation years
Two decades of full-stack engineering
Twenty-plus years across IT projects — including Prime Information & Communications — leading system design and build, and accumulating the depth that shows up in architecture decisions today.
- 16 May 2025
Yejua founded, and a turn as an online seller
Alongside registering the business, we sold online ourselves. Running into the real limits of global sourcing, margin and inventory led directly to building automation tools, and to launching yejuamall.com, an AI-assisted Japan sourcing mall we still operate.
- April 2026 – present
Into an AI agent ecosystem
Agent development began in earnest in April 2026. We built the AI Gateway on scalable infrastructure, released our own SaaS line-up, and delivered an advanced AI assessment simulation to an enterprise client.