AI products & internal agents
Two related things: AI products sold to customers, and agent platforms that automate a company's own internal work. We have shipped both, including an in-house Slack agent platform before off-the-shelf ones existed.
- 01Demos that impress and then cannot reach a single internal system
- 02Every team using a different model, with no shared or trusted interface
- 03Agents that need company data without being handed everything
- 04Cost per request that only becomes visible after launch
- 05No way to tell whether output quality is improving or drifting
Systems as tools
Internal services exposed over MCP or direct integration, so agents call real APIs instead of scraping screens.
Scoped data access
Agents read through views that exclude sensitive fields, enforcing the boundary in the tool layer rather than in the prompt.
Multi-tenant AI products
One backend serving several branded consumer products, with persistent memory, voice and media generation.
Retrieval that is grounded
RAG over your own documents, built so answers can be traced back to a source.
Evaluation
Test sets and scoring, so a prompt change is a measured decision rather than a hunch.
The gap between a demo and a system
Most AI work fails at the same place. The model is fine. The integration is missing. An assistant that cannot see your tickets, your orders or your documentation is a search box with better manners.
Everything below is built the other way round: start from the systems, expose them properly, then decide where a model helps and where a plain script is more reliable.
Internal agents
Sakurai is an in-house Slack-native agent platform built on Nest.js: more than fifteen focused agents covering support, product management, code review, release notes, moderation and analytics, with internal services (tickets, support, knowledge, metrics) exposed as MCP tools. It replaced people juggling three to five different models with a single named bot, and it was built before the off-the-shelf agent platforms were usable.
AI products
Companion AI is a multi-tenant AI-companion social platform on Cloudflare's edge: autonomous personas with persistent memory, real-time text and voice chat, AI media generation and a virtual-gift economy, with four live white-label tenants served from one backend across roughly 108k lines of TypeScript, localised into nine languages.
MandarinAI closes a loop between spaced repetition and reading: an FSRS scheduler decides which words you are about to forget, and a language model writes graded stories using exactly those words. Socialog exposes its whole product surface over MCP and REST, so an agent can schedule and publish without a human in the loop.
Where we will tell you not to
If a deterministic script does the job, we will write the script. Language models are worth their unpredictability when the input is genuinely open-ended. When it is not, they are an expensive way to make a reliable process unreliable.
Tell us the one thing it has to do.
If it's sharp enough to prove in a week, we'll say so and give you a date. If it isn't, we'll tell you that too, along with what we'd build first instead.