The starting point
AI is useful when it sits inside the work, not in a separate tab. A summary is worth more next to the email it summarises. An answer about a project is worth more inside the app where the project lives. We also wanted to avoid two traps we saw elsewhere: tying a product to one model provider, and trusting generated text because it sounds confident.
Lumi
Lumi is Lacspace’s AI assistant. It appears where people already are:
- On lacspace.com, Lumi answers questions about Lacspace and its products.
- In Lacspace One, Lumi is built into the app for questions about Lacspace and the work in progress.
- In Lacspace Mail, Lumi writes summaries and action items, drafts replies in your tone, translates and gives a pre-send review. It only sees the message or draft you are looking at, never the whole mailbox and never credentials. Drafts land in the composer for you; Lumi never sends on its own.
- In StockYatra, our trading simulator, Lumi AI is built in, with more of it on the Pro plan.
The toolkit behind it
We built Lumi from small, separate packages and published them on npm so anyone can use them.
- The AI Kit covers the basics of talking to a model: one chat and streaming interface across providers, typed prompt templates, token counting and budgets, repairing the JSON a model returns, splitting long text, reading streamed responses and defining tools a model can call.
- The AI App Kit builds on it for complete applications: embeddings, an in-memory vector store, retrieval-augmented generation, an agent loop that calls your own tools, conversation memory, evaluation scorers and a registry of model providers.
Every package has zero runtime dependencies, holds no keys and favours no provider. You bring your own key and choose the model, and you can change it later without rewriting your app.
Where it helps in a working day
In Lacspace Mail, the parts of Lumi that matter most are the quiet ones. Summaries and action items appear on long threads so a reader knows what is being asked before reading every reply, and they work even without an AI key. A draft written by Lumi lands in the composer in your tone for you to edit, and a pre-send review reads it once more before it goes. Translation helps teams who write to customers in more than one language. Each of these sits next to the message it is about, so nobody has to copy text into another tool.
In Lacspace One, Lumi sits beside the project: a client can ask about Lacspace and the work in progress from the same app where they approve releases and read invoices.
How we keep it honest
- Cheap checks first. A rule, a cache or a lexical screen runs before any model call, so the model only handles what is left.
- Checked against the source. Generated text is scored and compared with its source rather than trusted because it reads well.
- Visible limits. Our product pages say plainly that AI features can be wrong and should be checked before you rely on them.
What’s next
We keep adding Lumi to the places our clients work, and we keep publishing the parts that make it possible. Read more on the research page, try Lumi in Lacspace One, or browse the packages at developer.lacspace.com.
