Vela Request early access
For pharma analytics and forecasting

An analytics and forecasting team in software, for pharma.

Vela does the whole engagement, from auditing the client's data to the read-out, in hours or days instead of weeks. Every number is traced to its source and to the code that computed it.

Early access is by invitation, for expert consultancies and biotech teams.

The Vela workspace with an invented demo project open. Projects are on the left; in the middle, the project's steps (data audit, scorecard, forecast, test versus control, reach and overlap, read-out), each with a one-line result; on the right, the project's question, five problems found in the data and the questions to put to the client.
The Vela workspace on an invented demo project. Quillfeather and Quilvara are made-up names; no real data is shown.

The problem

Pharma decisions wait weeks for numbers no one can check.

Weeks of analyst time

A launch forecast or a program read-out from a consulting team takes weeks: cleaning the data, building the model, writing the deck.

Numbers that can't be traced

A figure in a slide rarely says which file, which formula or which assumption produced it. A chat model's answer says even less.

Accuracy no one has measured

Few forecasters can show how their method would have done on past launches, using only what was known at the time.

What it does

The whole engagement, in one conversation.

Open a project, add the client's files and ask the question. Vela does the work with its tools, shows each result beside the conversation, and suggests the next step.

  1. Data audit

    What is in each file, what is missing, what looks wrong, and the questions to put to the client.

  2. Descriptive work

    Scorecards against targets, trends, reach and overlap across channels.

  3. Launch forecasts

    Three-year US net sales for a drug launch, from several methods, each backtested on past launches.

  4. Test versus control

    Did the program work? The lift, with a range, and a warning when the groups were not alike to begin with.

  5. The read-out

    A client-ready summary written only from traced results, with every figure checked before it is shown.

  6. A link the client can check

    A read-only page where every figure opens to show the data and the calculation behind it.

A test-versus-control result on invented data: a lift of 11.6% with a 95% interval of 10.8% to 12.4%, 70 test and 90 control accounts, and a 'Read this first' warning that the two groups differed before the program.
Test versus control: the lift, its range, and a warning the analyst reads first.
A read-out on invented data. Each figure in the summary carries a small numbered reference to the recorded result it came from, followed by a table of the files that were examined and the problems found in each.
The read-out: each small number points to the recorded result behind that figure.

Why Vela

Built so a client can check every number.

A general AI model can run a forecast on data it is handed. Vela holds the whole project, and it is built so the work stands up to review.

Every number traced

Each figure links to its source data and to the code, and the code version, that computed it.

Code does the arithmetic

The AI gathers inputs, runs tools and writes. It never does the math, and a write-up whose numbers don't match a recorded result is refused.

Backtested without hindsight

Forecasting methods are tested on real past launches, each using only what was known before its launch date.

Bring your own method

Experts teach Vela their method in the Method Builder: an interview, then a Method Card, then the build and a backtest on past launches. The method stays private to its author unless they choose to publish it.

Client data stays in its workspace

Each client's data lives only in that client's private workspace. It never trains shared models or methods, and separation is enforced by the system's design, not by the AI's judgment.

Proof, so far

Measured on real past launches, and stated plainly.

In a walk-forward backtest, each past launch is forecast using only information available before it launched, then compared with the sales that followed.

Backtest of 2026-11-16

Median miss on year-3 US net sales (lower is better)

59 past US drug launches with separately reported sales. The naive baseline is the simple forecast any method must beat; a method ships only if it does. Figures change as the dataset grows.

What this does and doesn't show

  • The best method beats the naive baseline on real launches, without hindsight.
  • New forecasts are locked with a timestamp and a hash, and scored as real sales arrive.
  • The launches are those whose companies report US sales separately, about one in six approvals, so results may not carry over to every launch.
  • A 55% median miss is still large. We don't yet claim a head-to-head win over analyst forecasts.

Who it's for

For the people who do this work today.

Expert forecasting and analytics consultancies

Small firms and independent consultants who forecast and analyse for pharma clients. Take on more projects with your own methods, and hand clients work they can check line by line.

Biotech commercial and business-development teams

A fast, defensible second opinion before licensing a drug or planning a launch, without a multi-week consulting project.

Team

The founders.

Justin

Co-founder & CEO: product and engineering

Gavin

Co-founder: data and customers

Bring us a past launch or a live question.

We are inviting a small number of expert consultancies and biotech teams to try Vela on their own kind of project.

Request early access