on a mission to free the gluemans. founder & cto @ wordware
San Francisco, CA
Joined February 2022
- Tweets706
- Following745
- Followers2.2K
- Likes2.6K
ok, @kozerafilip cooked, this is the most beautiful blog post since @kitlangton's opencode 2.0 blog post
sauna.ai/blog/sauna-2
got a little carried away and turned pagedb into a vector database too
semantic search over all 4.85m hacker news stories. pinecone would charge ~$343/month for this, turbopuffer $16
one durable object: ~$5 🧵
code and every benchmark:
github.com/nyacg/pagedb
the best thing we did for our agent was take its tools away. all but one.
mcp: hundreds of tool manuals in context, one round trip per page of results, and you pay for every one.
code mode: one tool that runs code. the agent writes the script, and only the answer comes back.
code: github.com/nyacg/pagedb
what it can’t do (yet):
- apps only run queries defined up front, each with an index. a new one means a rebuild
- the 100gb is built offline. new writes stay in the object (10gb cap), folding them into R2 is next
- one object = one thread in one region, no replicas
and the numbers:
- $9 is a 30s load test (1,000 sockets from my laptop, all watching one list) scaled to a month
- competitor prices are my list-price estimates, every write sent to every user. with lighter fan-out they’re $190–2,100
all of it: github.com/nyacg/pagedb#limi…
inspired by @tantaman's r2db, which proved this was possible at all:
nitter.cf/tantaman/status/211094…
i rebuilt it from scratch with claude code: a rust engine compiled to wasm, the sync layer and four live apps
How fast and slim can we make Zero's IVM? To find out, we looked for the tiniest and most constrained runtime to put it in. Here it is. Zero's next-gen IVM engine in a single DO. Querying a 100GiB DB stored in R2 and maintaining live queries.
demo.r2db.dev/
