Two-week pilot

One number, moved in two weeks, or you pay nothing.

DPR Labs is the senior AI studio for teams tired of demos, dashboards, and rented black boxes. We agree on one business number before we start. We move it on your real data, prove the move with a frozen check, then hand you the system.

Most AI work is slop with a nicer deck: a chatbot nobody asked for, a private dashboard nobody can audit, a model score that falls apart when the messy inputs arrive. Our gold mascot is not a hype animal. It is a warning label. If the work cannot survive measurement, ownership, and handoff, it is not elite work.

Book the 20-minute scoping call

See the method first

The contract

Steal this sentence before you buy any AI pilot.

Every pilot starts with a metric contract. Not a strategy doc. Not “make us AI-first.” One sentence that can be read by finance, operations, and the person who will inherit the work. If the sentence sounds vague, the pilot is not ready.

Template: We will move ___ from ___ to ___ within two weeks, measured on ___, checked by ___.

Example: We will move support-ticket first-response resolution rate from 41% to 60% within two weeks, measured on a frozen holdout of 500 real tickets, checked by your ops lead running the eval suite we hand over.

The number is agreed in writing and does not change mid-pilot. It has to be something you already measure or can measure without theater. No moving goalposts. No basket of vanity stats. No blended “quality score” that lets everyone argue after the invoice.

Metric contract builder

Write the scoreboard before the model exists.

Use this in the scoping call. If the blanks are hard to fill, that is useful information. A weak metric is not a small detail. It is the crack where the whole pilot leaks out.

We will move from to within two weeks, measured on , checked by .

We will move metric from today to target within two weeks, measured on real data, checked by your person.

Ownership

You own the system, not a login.

When the two weeks end, you own the whole thing. We do not hide the useful parts behind our account, our runtime, or our private prompts. If a vendor will not hand over the pieces, you are not buying an AI asset. You are renting a dependency with better branding.

Source code

Every line we wrote, in your repo, under your license, with no obfuscation and no mystery binary you cannot rebuild.

Weights and adapters

Any fine-tuned weights, LoRA adapters, and exact base-model references needed to re-serve the work without us.

Eval suite

The scoring harness and frozen test set that decides whether the number moved, so your team can rerun the verdict.

Prompts and config

All prompts, system messages, decoding settings, and parameter files, versioned with the reason each choice exists.

Deploy scripts

Containers or infrastructure scripts that stand the system up from zero in your accounts, not ours.

Data pipeline

The ingest, cleaning, and formatting code that turns raw operational data into what the model sees.

Runbook

A plain-English operating guide: how it works, how to refresh it, how to roll back, and what to inspect when the number drifts.

Honest limits

What two weeks cannot buy.

A two-week pilot can prove and hand over one owned improvement. It cannot make your company “AI-native.” That phrase mostly means nobody has agreed who owns the work after the demo.

  • It cannot fix more than one metric. A second number is a second pilot.
  • It cannot rebuild broken data. If the data to move the number does not exist, we decline the pilot.
  • It cannot deliver a fully hardened, on-call platform with a 99.99% service promise.
  • It cannot guarantee the gain holds forever. Models drift, workflows change, and the eval suite is how you catch it.
  • It cannot pass your formal compliance, security, or legal audit inside the same window.
  • It cannot move a number nobody can measure. No honest scoreboard, no pilot.

This is the controversial part: saying no is part of the product. We would rather reject a bad pilot in twenty minutes than sell you a beautiful failure.

ROI calculator

Check whether the bet is even worth making.

Not every workflow deserves a model. If the math cannot pay back a small pilot, the right answer may be a checklist, a script, or leaving the work alone.

Payback
Annual saving

This calculator runs only in your browser. Nothing is sent anywhere.

Price

Fixed price. No move, no pay.

Most pilots land between $8,000 and $15,000, depending on how much data plumbing the one number needs. No hourly meter. No surprise line items. No scope-creep invoice after everyone is already emotionally committed.

If the agreed number does not reach the agreed target on the agreed check within two weeks, you owe nothing and keep every artifact produced along the way. The refund is real because the scoreboard is set before we take the work, and your data checks it. We only accept pilots where the number looks movable. We are betting our time against your outcome.

20-minute scoping call

The call decides if there is a real pilot.

You leave with a fixed price and start date, or an honest no. Both are useful. The agenda is intentionally narrow.

0-3 minutes

Name the one number and why it matters commercially.

3-8 minutes

Set the baseline, target, and current measurement path.

8-13 minutes

Inspect the data source, access path, and whether it is clean enough for two weeks.

13-17 minutes

Define the frozen check that decides pass or fail.

17-20 minutes

Choose go or no-go, fixed price, start date, and owner.

Close

Book the 20-minute scoping call.

Bring one ugly workflow and one number you care about. We will tell you if AI can move it, how we would prove it, what you will own, and what we refuse to pretend.

Book the scoping call

Read the pilot autopsies