Applied AI
AI features that ship, measured against what they actually cost.
Most AI projects fail in the same place: a convincing demo that nobody can put in front of customers, because nobody measured whether it is right often enough, or what it costs at volume.
- From
- $6,500
- Typical timeline
- 3–6 weeks
- Stack
- Python · TypeScript · PostgreSQL · Vector search · Anthropic API
We build the boring parts. Evaluation sets so you know the accuracy. A cost model so you know the bill at ten thousand queries. Fallbacks so a model outage is not an outage for your product.
What this covers
Search and Q&A over your own content
Ask questions of your documents, policies, or product data and get answers with citations back to the source. The single most requested AI feature, and the one most often built badly.
Structured extraction
Turning PDFs, emails and forms into clean, validated data. Invoices, applications, reports — the work that currently occupies somebody's whole afternoon.
Triage and routing
Classifying and routing incoming support requests, applications, or documents so the right person sees the right thing first.
Workflow automation
Multi-step processes with a model in the loop, built with explicit checkpoints so a person stays in control of anything consequential.
What you get
- A shipped feature inside your product, not a prototype
- An evaluation set and measured accuracy against it
- A cost model at your projected volume
- Fallback behaviour for model outages and rate limits
- A written note on what the system cannot do
Stack
- Python
- TypeScript
- PostgreSQL
- Vector search
- Anthropic API
- OpenAI API
- Docker
We use a small, well-understood stack deliberately. Fewer surprises, faster decisions, and estimates that hold.
Packages
AI Feature Sprint
A product team adding its first AI capability.
$6,500
3 weeks
Common questions
- Will our data be used to train a model?
- No. We use commercial API tiers that do not train on submitted data, and we will confirm the specific terms of whichever provider we use in writing before any of your data touches it.
- How do we know it is accurate enough?
- We build an evaluation set from your real cases before building the feature, then measure against it. You get a number, not an impression — and if that number is not good enough, we say so.
- What does it cost to run?
- We give you a cost model at your projected volume before you commit to the build. Running cost is a design constraint from day one, not a surprise on your first invoice.