what i build
every llm provider rate-limits or goes down eventually, and apps built
on one provider stop with it. flippy is an inference router that
fails over across four providers automatically, plus a unified hub
for chat, rag, tool-calling, vision, tts and stt.
what i learned: failover is easy to write and hard to test — the value is in simulating provider failures before they happen in production.
refreshes every 2 hours, live in production
job boards serve stale postings, so searchers waste hours applying to
filled roles. just-hired shows only direct-employer postings from the
last 12 hours, fetched by a cloudflare worker (which gets past job
bank's datacenter-ip block where github actions couldn't) and
redeployed automatically.
what i learned: egress location matters more than code quality when a source blocks datacenter ips — pick the runtime the source trusts.
44 parser tests · 30-day live-mode eval gated behind a flag
an llm voting real money on its own judgment is one bad prompt away
from losing it. this is a paper-mode crypto signal engine: the
pipeline computes every metric first, then a free-tier llm casts a
structured buy/skip vote per token — the model judges inputs it
can't fabricate, and the engine can override it.
what i learned: keep the model as judge, never as source of truth — and gate anything risky behind an explicit opt-in flag.
hard send caps + opt-out enforcement in code
cold outreach usually claims too much and proves too little. this
system finds one measurable defect per prospect, prices the fix, and
drafts one email at a time with evidence the recipient can verify in
twenty seconds. send limits and unsubscribe handling are enforced by
the pipeline, not by discipline.
what i learned: safety rules that live in config get ignored; rules enforced in code survive pressure.