Made some animations and loaders for @aceternitylabs@mannupaaji 😋
(last post glitched)
View at: https://t.co/zOuQACZBcW
Would love your feedback and reviews! 🥰
the fix:
ground it on the current policy doc (synced daily)
verifier checks every answer against the source
confidence < 0.75 or unsupported claim → human reviews before sending
track freshness, groundedness, context recall.
“bigger model?” doesn’t fix stale data.
gave an AI eng interview for a remote company.
got asked this:
Your support bot says annual plans get a refund within 45 days. Actual policy: 14 days, updated 6 weeks ago.
the team measures: p95 820ms. errors 0.02%. uptime 99.98%. no alerts.
What went wrong?
your answer? 👇
nothing was “down.” the bot was pulling the old policy doc and answering from it.
DBs and queues fail loud: timeouts, 500s, crashes.
LLMs fail quietly. the response is fluent, fast and wrong.
latency and uptime only tell you the API answered, not whether the answer was right.
got shortlisted, had a call with the founder and eventually decided to pass, as the role was more aligned with content and strategy than the tech side im looking for.
happy to see the growth tho.