AI system design studies
How I'd Build It
Eight operational problems, each told the same way: how a team worked before, where it broke, one item followed through the system I would build, and then the full technical design. The organizations are hypothetical, written to make each problem concrete, and every figure is illustrative. None of this describes client work.
Start with a problem your team recognizes.
- Revenue operations Reliable lead intake A prospect waited two weeks because one automation run failed silently. The fix: record every submission before doing anything slow with it.
- Customer support Support drafts from current policy An assistant quoted a well-matched refund article that billing had replaced two weeks earlier. Relevance is not eligibility.
- Finance operations Invoice checks before ERP entry Automating the typing is the easy half. The dangerous half is the judgement that used to happen in the same motion.
- Reliability engineering Incident triage with linked evidence The signal was on a dashboard the whole time. Nobody looked, because the first plausible story filled the room.
- Software engineering Code review matched to change risk When every pull request gets the same AI review, nothing gets a careful one, and people stop reading.
- Legal operations Contract review across linked clauses Search found the liability cap and quoted it perfectly. It didn't follow the exclusion, the defined term or the amendment.
- Marketplace trust and safety A small model for listing review Everyone had an opinion about whether a small model would do. The design starts by replacing the opinions with a comparison.
- Sales compliance Call review inside a private network A 2% listening sample missed one adviser's pattern for months. Checking every call is easy in the cloud and not allowed.
Have a similar process?
Tell me where the work gets delayed or needs checking. We can use one example to decide whether a custom system would help.