product AI that ships

saas founders sitting on a model poc with no product surface — no auth path, no empty state, no way a user actually finishes the job.

// not for
rfp theater · pure research · demos that never meet a user

• one instrumented feature path in the product, not a sandbox
• fail modes designed: slow, wrong, empty, rate-limited
• you own the repo when we leave

• one scoped feature path (named job-to-be-done)
• model/tooling layer: provider wiring, structured output or tool calls
• backend path: server boundary, auth check, rate limit, audit hook
• customer UI: primary path + empty / error / loading / partial states
• admin or ops control: flag, config, or thin authenticated panel
• instrumentation: cost/latency read, enough to debug at 3am
• eval seed: golden examples for the path
• handoff: repo, env map, short run notes

// out

• multi-workflow agent fleet (→ agents)
• full corpus retrieval product (→ rag)
• production eval CI gate as the whole engagement (→ evals)
• pure visual rebrand with no model path
• model research / fine-tune vanity with no user job
• unlimited feature backlog under one band

1. 01

   tell us the job

   one path, who runs it, what done looks like.

2. 02

   sample the path

   free sample or paid spike — something running.

3. 03

   build in a room you can watch

   model, UI, admin for that path.

4. 04

   ship + keys

   you own it. 90 days if something breaks.

Next.js · Vercel AI SDK · your model host

named tools. no stack theater. if it isn't on the critical path for this line, it isn't listed.

from $18k · 3–6 weeks

one feature path, instrumented

bands filter the room. exact number after a call — not a 40-page sow.

we already have a model POC. what do you actually add?

the product path. auth, states, operator controls, instrumentation, a handoff you can run without us. the model call is the easy part once. the path is what users touch every day.

is this agents, rag, or something else?

product-ai is one feature path inside your product. multi-step tools with human gates → agents. answers only from your corpus with citations or silence → rag. screen trust only → design-ship. shipped a prompt and prayed → evals. book if you're between two.

do you bring the model or use ours?

either. we wire your host or stand up a sane default for the slice. we don't sell a proprietary model. we sell the path around it.

how do you price this?

from $18k · 3–6 weeks for one feature path. exact number after a call. no 40-page sow theater.

book if you're not sure which. these two sit next to this line.