About us
QAI Lab is an applied AI research lab.
We work on the four things an agent needs to be useful on real data — retrieval, memory, tool use and evaluation — and our research leaves as production software rather than as papers. QAI Finance, QAI Earth and QAI Insights are where it is tested against people who need an answer rather than a benchmark.
Why a research lab ships infrastructure
Agentic research stops being research the moment it touches real data. A question about a company's margins or a river basin's discharge needs filings, map layers, time series and documents in one place, under one permission model, with one audit trail. That is where most agent projects die, and it is not a paper — it is a schema, a runtime and a bill.
So we built the layer underneath, and then we had to operate it. A metering bug shows up in our own billing before a customer's; an outage takes our own products down first. That is the whole argument for a lab shipping infrastructure: the research is only tested once something depends on it.
The position we are taking, which could turn out wrong: the hard part of agentic systems is no longer the model. It is everything that decides what an agent may touch, what it remembers, and what it costs — and that has to be settled below the agent, not inside its prompt.
What we have not solved: evaluation. We can tell you what an agent did and what it cost, and we can bound the damage when it is wrong. Deciding whether its judgement was good, at the scale these systems now run, is still an open problem — ours and everyone's.
Our Values
How we decide what to build, and what to believe.
One data plane
New data types go into the same Postgres, under the same permissions, before they go anywhere else. Four datastores that have to be kept in sync is a class of bug we would rather not ship.
Evidence, re-checked
Anything we claim has a table, a migration and a call site behind it. When a source moves, the judgement resting on it gets re-derived rather than quietly left standing.
Metered, not estimated
Model usage is accounted per message, per product and per organization. Usage that cannot be attributed does not count as shipped.
We run on it first
Our own products are the only things on this layer — same schema, same runtime, same quotas. A metering bug shows up in our own billing, and an outage takes our own products down first.