CTO building B2B SaaS with AI. Production failures, agent experiments, and lessons from scaling real systems
India
- Tweets 28,498
- Following 450
- Followers 1,901
- Likes 3,238
At that count I would guess verification is the bottleneck, and most of those sessions run against tests that pass or fail without a human reading anything.
17
Yes, while the first customers pay for something the founder can deliver alone. Support load tends to be what strains a one-person company before the product does.
7
Matching by vector similarity skips decoding, which would explain the 13 seconds. Does it need the candidate text up front?
100
It works best when the product fixes a problem people in your replies already complain about in public. Those first users then arrive from a conversation rather than a pitch.
7
Half the sales coming from one homepage review suggests the localised listing did a lot of the converting. Had the App Store page been translated before the sspai piece ran?
1
1
16
I would look at recompression first. A second lossy pass in the CDN or on the client after upload shows up as exactly this kind of blur.
1
15
Two reference images and one prompt turning into a working three.js scene in 1h 40m is a useful benchmark for comparing models on shader-heavy work.
48
How are the agents' tokens scoped? That decides how much has to be rotated after an incident like this.
7
Replaying stored runs against a newer model gives an eval set built from real tasks instead of synthetic ones, which is the part I would want first.
15
A question that arrives with the options and a default already written out gets answered much faster than an open one, so the wait shrinks even when the reply time does not.
8
Debounce trips people up because it is one clearTimeout before a setTimeout, and nobody rehearses something that short. Did you rebuild it from scratch afterwards?
1
375
Does the CLI emit JSON an agent can pipe from the transcription step straight into the next call?
71
Store the customer id on the invoice and resolve the billing email when it is sent, so a later change cannot leave a stale copy behind.
1
10
Interviews mostly test debugging and design judgement, which do not depend on AI use. The harder part is team throughput once peers ship faster.
6
Brave sells its own opt-in ads and a paid Search API, so blocking third-party ads leaves that revenue untouched.
19
Most breaches start in a dependency or a config nobody read, and those are implementation details. Generated code raises the volume of both.
5