Ship the AI feature.
Skip the demo.
Most AI work dies at the demo. The model talks, the room nods, and three months later the feature still isn't in front of users. Here's the playbook we use to get it out the door.
- Read
- 9 min
- Last updated
- May 2026
- By
- Eli Park, founding partner
- Stage
- Series A – C product teams
The half-built AI feature is the most common artifact of 2025. A retrieval prototype in a fork of the repo. A demo that worked twice. An eval suite someone meant to write. It sits in a Slack channel called #ai-experiments, and the people who could finish it are the same people shipping the rest of the roadmap.
We don't do greenfield AI. We finish the work you started. That's the entire offer, and it's deliberately narrow. We come in for 6–14 weeks, take the half-prototype the smart person on your team built between meetings, and turn it into the version that ships — wired into the product, instrumented, on-call, documented.
§ 01 — The thesisFinishing is the hard part.
Anyone can get a model to do the thing in a notebook. The hard part is the surrounding 2,000 lines: the retrieval that doesn't hallucinate, the evals that catch regressions, the cost ceiling, the fallback for when the API is down, the prompt versioning, the analytics, the rate limits, the security review. None of that is glamorous. All of it is why your feature isn't out yet.
“We don't do greenfield. We finish the work you started. ”
We're opinionated about the surrounding 2,000 lines because we've written them thirty-seven times. We have a default retrieval stack, a default eval harness, a default observability layer. You're free to disagree with any of them — and clients often do — but we never start from zero, which is the difference between a six-week engagement and a six-month one.
- ✓The feature, shipped to prodNot a demo branch. Behind a flag, on real traffic, with rollback.
- ✓An eval suite that runs on every PRGolden set, regression tests, hallucination + cost budgets.
- ✓Observability hooked upPer-request tracing, prompt versioning, cost-per-user dashboards.
- ✓An on-call runbookFor when the model API has a bad afternoon. Yours to keep.
- ✓A handover weekWe pair with your team for five days at the end. No black boxes.
- ✓Office hours for 90 daysWeekly call, async Slack, no extra invoice.
§ 03 — How it goesFour phases, one quarter.
The default shape is 12 weeks, broken into four phases. We can compress to 6 if the prototype is far along, or extend to 14 if there's a meaningful infra rewrite hiding underneath. We won't extend past 14 — if it needs more than that, what you actually need is a hire, and we'll tell you so.
The audit is non-refundable, but it's also non-binding: you can read the diagnosis and walk away. Two clients have done that, and we considered the audit a success in both cases. Knowing what's actually broken is worth the fee even if you fix it yourself.
§ 05 — The catchWhat this costs you.
Engagements run $90k–$240k depending on scope, paid in three milestones. The audit is $12k and rolls into the engagement fee if you continue. We work with one client at a time per engineer — we're a team of four, so we run two or three engagements in parallel, never more. If we're full, we'll tell you, and we'll tell you who else is good.
Skip ahead and start a conversation. Most engagements begin with a 30-minute call where we figure out together whether you're in the “we can help” zone or the “you don't need us” zone. We're roughly 50/50 on that split, and we're proud of it.