Prospect Scoring Engine
Raptor reads more than 160 signals about every prospect and learns, from your own results, which ones matter. A small recurrent neural network follows each prospect's history of touches and estimates the chance of a positive reply, a meeting or a win.
Also on Mac (Apple Silicon, beta): Download · Install guide
A score nobody can explain is just a number.Raptor shows what moved each estimate, and says plainly whether it is using starting weights or has learned from your results.
Step by step.
Read the signals
167 signals in nine groups: fit with your company, buying triggers, size and maturity, web presence, technology, needs in their own words, industry, news, and how reachable the right person is. A signal Raptor has no data for counts as zero, never a guess.
Add the history
A recurrent network (a GRU) reads what has happened to each prospect in order: pitches, follow-ups, replies, bounces and meetings, with the channel, the day and the spacing between them.
Estimate the chance
One output gives the chance of a positive reply, meeting or win. Autopilot works through the best-estimated prospects first, while your fit threshold still decides who is pitched at all.
Learn from your outcomes
The network trains on your own computer, only on your own outcomes. It needs about 40 prospects with an answer, at least 8 of which converted, and then retrains as new outcomes arrive.
Prove itself before it counts
A trained model is switched on only if, on prospects it never trained on, it separates the ones that converted from the ones that did not better than chance, with 95 percent confidence. Otherwise it stays off.
Built in.
167 signals, nine groups
Fit, triggers, size, web, technology, needs, industry, news and contact quality, each a named number you can inspect.
Why this estimate
Open any prospect to see the signals that moved its estimate up or down.
Honest labels
Every estimate says whether it comes from starting weights or from a model learned from your results.
Trained on your machine
PyTorch runs locally and trains in seconds. Your outcomes never leave your computer for this.
Honest expectations: until you have enough answered prospects, estimates come from starting weights that reflect what usually matters in B2B outreach. They are a sensible guess, labelled as one. Expect the learned model after weeks of use and a few dozen answered prospects, not on day one.
- Signals
- 167 in nine groups
- Model
- GRU over each prospect's history, joined with the signal vector
- Training data
- Your own outcomes only, on your computer
- Switch-on rule
- 40+ answered prospects, 8+ conversions, held-out AUC above chance at 95% confidence
- Retraining
- Automatic as new outcomes arrive
See why a prospect ranks where it does.
Run a hunt and open any prospect to see the signals behind its estimate.