We built Dreambase after spending four years watching the sync job become the most failure-prone part of every LLM application.
Founded in 2021 · Cambridge, MA · four engineers · bootstrapped.
Why we built this
Before starting Dreambase, Andrew Keil spent four years building ML infrastructure at a database software company in the Boston area. By 2020, most of the teams he worked with were integrating their first LLM features. Every one of them landed on the same architecture: structured data in Postgres, embeddings in a separate vector store, and a cron job or Celery worker keeping them roughly in sync.
"The sync job was the source of more on-call pages than anything else," Andrew says. "Not query latency, not embedding quality — the plumbing. It would fall behind during high write periods, skip records silently after a schema migration, or drift for hours after a deployment rollback. Users would edit a document and get stale retrieval for the rest of the day. The teams knew it was happening. They just had no better option."
In late 2020, he started asking a different question: what if the vector index wasn't a separate store at all, but an index on the table — just like a B-tree, but for approximate nearest neighbors? What if the query planner had visibility into both access paths and chose the cheaper one? The first prototype ran in early 2021. The core design held. The managed cloud launched in January 2025.
Dreambase is not a pivot or a side project. It is what we have been building for four years. We are four engineers. [email protected] reaches a founder.
The people building Dreambase
How we work
We document trade-offs, not just benefits. The concepts page says explicitly: if your workload needs above 50K writes per second or above 4096 dimensions, a dedicated vector store is probably the better choice. We are not trying to serve every database use case — we are trying to solve the dual-store problem for the teams it actually affects.
Every feature ships with a benchmark against real row counts and real embedding dimensions. The SQL + Vector guide publishes p50 and p99 latency numbers at 100K, 1M, and 10M rows. If a release regresses latency, the changelog says so — not just the release notes for the fix.
We are four engineers building one database. We do not have a sales team, a developer relations org, or a growth function. [email protected] reaches Andrew directly. We have been bootstrapped since 2021 — every decision is made by the people writing the code, not a board slide.