Service 04 · LLM features · Document understanding · Assistants
AI features that survive contact with users.
Document understanding, assistants and search built on top of LLMs, with an evaluation set, cost controls and a source for every answer. The same pipeline we run in Docsee.
Pipeline · as shipped in Docsee● running
01 — What we build
Four things that actually ship.
Not "AI strategy". Features with a user, a metric and a cost per call.
02 — How we keep it honest
A demo is easy. A feature that is right 98% of the time is the job.
Every project starts with an evaluation set: real inputs from your data and the answers you consider correct. Every change to a prompt, model or retrieval step runs against it, so quality is a number, not a feeling.
Every answer carries its source. Cost per call is tracked from the first prototype, because a feature that costs more than it earns does not ship.
Evaluation
A test set from your real data, run on every change
Sources
Every answer points to where it came from
Cost
Caching, routing to smaller models, a budget per feature
Privacy
On-device where possible, no training on your data, written retention rules
03 — How it runs
Prototype first. Then decide.
Week 1
Use case and data
You get: the one feature worth building first, what data it needs, and a go or no-go with reasons.
Weeks 2–4
Prototype and evaluate
You get: a working prototype on your data, an evaluation set, a quality number and a cost per call.
Weeks 5–8
Integrate
You get: the feature in your product, behind your auth, with monitoring, limits and a fallback when the model is down.
Ongoing
Watch and improve
You get: a monthly quality and cost report, model upgrades when they pay off, and the evaluation set as your asset.
04 — Proof
Case 02 · Docsee · Live
Scan any document. Then just ask.
Document understanding with sources, shipped to the App Store as our own product. Read how the evaluation set and cost controls were built.
Read the case →
Pricing
4 weeks, fixed price, ends with a working prototype, an evaluation report and a go or no-go. Integration estimated separately on what the prototype shows.
Model costs billed at provider price, no markup.
05 — Questions
About AI projects.
Whichever passes your evaluation set at the lowest cost. We have shipped on Claude, GPT and open models, and routing between them is part of the build, not a rewrite.
Provider APIs with no training on your data, on-device processing where it fits, and retention rules written down before the first call. For regulated data we can run open models in your cloud.
Three ways: every answer carries its source, the model is told to say "not found" rather than guess, and the evaluation set measures how often that still fails. You see the number before launch.
Cost per call is tracked from the prototype onward. Caching and routing to smaller models usually bring it down several times over; the prototype report shows the figure for your usage.
Book a scope call
Bring the data. We bring the honest answer.
Thirty minutes with a sample of your documents or tickets. If an LLM is the wrong tool, we say so on that call.