← All posts
AI without the theatre
Most of the AI work we do is unglamorous. Document extraction, routing, summarisation. Here’s what we’ve learned shipping it.
Roshan Soni · Founder · Engineer
The most useful AI projects we’ve shipped lately look nothing like the demos. They’re intake assistants reading PDFs. They’re chatbots answering policy questions from a private knowledge base. They’re classifiers turning noisy free-text into structured rows. They’re boring. They work.
What we look for
- A task someone does today that costs hours per day.
- A clear success metric — ideally something already tracked.
- A boring failure mode (human takes over) instead of a confident wrong answer.
- Private data we can ground the model in.
What we avoid
- “Chat with our website” projects with no operational ask behind them.
- Models hallucinating into production without a fallback.
- Pilots that have no measurable owner.
If your AI idea sounds like a press release, that’s a hint. If it sounds like a tired ops team’s wish list, that’s usually the start of something that ships.