masato

something i didn't expect from doing NLP research: evaluation methodology became my default product habit. 30 annotated test examples before any feature. felt like overkill in grad school. now it's how i ship without bugs.

day 37. start of may. april ended at 36 X followers, 8 here. started both at single digits. 6x on the top of the funnel. 0x on revenue. posting works for attention. nothing else. the bottom of the funnel is where may goes. #buildinpublic

day 35. last day of april. month in numbers: 6 → ~50 followers 3 products live 0 paying users 2 weeks of consistent posting building was easier than expected. distribution still hasn't clicked. #buildinpublic

day 34 of building publicly. 3 products live. 0 paying users. working on the focused one this week. freelance work picked up too. balance is the actual job. #buildinpublic

took a week away from building. came back to 4 products at 80% complete. distance helps you see which one actually matters. the answer was not the one i expected. #buildinpublic

something i've been noticing in my own work: the weeks i shipped 1 concrete thing moved more than the weeks i talked about shipping every day. closing that gap. #buildinpublic

claude code session that worked today: spec written in markdown first. 20 minutes focused coding. done. vs sessions where i "just start typing": 2 hours, mostly fixing what i broke. short sessions, clear contracts.

something i'm reminding myself: my stack is irrelevant until someone pays for the output. stack is leverage. revenue is the test. still working on the second part. #buildinpublic

solo SaaS with $0 revenue: shipping the product was 30% of the work. finding users who care is the other 70%. i'm only realizing this 4 products in.

3 days on bluesky. 5 followers, 6 posts. different platform, different reset. posting what i'm actually doing here too. no expectations. small base, quiet feed. the early days here feel like 2010 twitter.

my actual stack (3 products in production, 0 paying users): next.js supabase claude api stripe (configured) vercel resend monthly cost: ~$20. revenue: $0. build is solved. distribution isn't.

sitting with 4 products today: kashite (lending tracker) yomu (AI doc reader) phrasely (rewrite tool) sabaku (focus app) picking 1 by tonight. the others go to a paused folder. closing tabs is harder than opening them. #buildinpublic

pgvector vs pinecone real-world choice: under 1M vectors → pgvector over 10M vectors → pinecone between → depends on query patterns most "AI startups" have under 100K vectors. you don't need a vector database. you need a database with vectors. #ai #postgres #rag

most "build in public" reveals: what's working: announced loudly what's not: silence what almost killed the company: never tweeted the curated honesty is still curation.

the "I'll just rebuild it" trap: every time you decide to rewrite, you're making a bet that the new version will reach feature parity faster than the old one accumulates new bugs. that bet rarely pays off.

supabase RLS pattern that scales: don't write 50 policies per table. write 1 policy per access pattern. policy: "users see their own rows" applies to read, write, update, delete across most tables. abstract early. RLS is code. #supabase #postgres

stripe webhook gotcha that bit me twice: your handler must be idempotent. stripe retries on any non-2xx response. if your db write succeeds but you return 500, you'll process the same payment 3 times. always check event.id before writing. #stripe #webdev