An always-on scout that hunts for dream customers and dream partners on repeat, learns your taste from every yes/no you give it, and hands the best-fit few to Hundo to work the relationship.
You cannot write a perfect filter for "dream partner." You know one when you see one, but the definition lives in your gut, not in a spreadsheet of criteria. A self-improving loop solves that by flipping the job: instead of you describing the target up front, the machine learns your taste by watching you judge real candidates, and gets sharper every single round. Here is the whole engine in seven steps.
Hand it a handful of your best current partners (or customers) as gold-standard examples, plus a plain-language description of who you want. No perfect criteria needed. Five good examples beats a long list of rules.
Every candidate gets turned into a numeric fingerprint of who they are, built from their firmographic and profile text. Similar businesses land near each other, so "looks like the ones you loved" becomes something the machine can measure.
A lightweight model predicts a dream-fit score for each candidate from its fingerprint, and, just as important, how confident it is in that guess.
It hands you a small batch to react to. Not random: a deliberate mix of high-confidence yeses to act on now, and a few it is genuinely unsure about (more on that below).this is the clever part
You thumb up or down (an LLM can pre-filter the obvious junk first so you only see plausible ones). Every judgment becomes a training label. This is the only work you do, and it takes seconds.
The model retrains on everything you have ever judged. Your taste sharpens into the score. Round two is smarter than round one, round ten is scary-good.
Only high-confidence picks pass the spend gate, get a paid email lookup, and route into Hundo (partners) or Instantly (customers). Then the pool refreshes and it loops back to step 2, on repeat.
Every yes/no feeds back into step 6. The loop never stops learning, and it never stops hunting.
A dumb version would only ever show you what it already thinks is a yes. That fills your pipeline but the machine never gets smarter. The loop balances two modes on purpose:
When it lacks signal, it surfaces the candidate it is most torn about. Your answer there teaches it the most, fast. This is how it maps the fuzzy edges of your taste instead of just guessing safe.
Once it has enough signal, it surfaces the candidates most likely to be a yes, so real dream-fit leads flow into Hundo now. It tilts here more and more as it learns.
Early on it explores to learn your taste. As confidence builds it exploits to fill your pipeline. You get both a smarter model and a steady stream of real prospects out of the same motion.
The loop does not care whether "fit" means buys from you or refers to you. Swap the seed examples and the judge question, and the identical engine runs a Dream Customer radar or a Dream Partner radar side by side. That is the whole point you spotted: ICP and IPP are the same machine, pointed at different gold examples.
| ICP loop · dream customers | IPP loop · dream partners | |
|---|---|---|
| Seed with | Your best current clients | Your best current partners (the ones who actually send deals) |
| "Fit" means | Would buy, would be a great client to serve | Serves your buyers, complementary not competing, has reach and a reason to refer |
| Judge question | "Would this be a dream customer?" | "Would this be a dream partner: right audience, aligned, and motivated?" |
| Signals it weighs | Firmographics, pain fit, budget shape | Audience overlap, complementarity, credibility, referral incentive |
| Hands off to | Instantly (volume) or your sales motion | Hundo, for the relationship-first recruit |
Why IPP is the sharper play: a dream partner is not "someone who looks like my customer," it is "someone who serves my customer." That is a subtler judgment and exactly the kind of fuzzy target a learned loop beats a hand-written filter at. Your gut already knows it. The loop just bottles your gut.
No paid email lookup, no enrichment, no outreach until the model is confident and you have capacity to act. The radar can scan thousands for free and only ever spends a credit on a lead it is sure about. Cost scales with quality, not with volume.
OpenOutreach uses heavy academic math (a Gaussian Process with Bayesian active learning). At your scale that is overkill. The loop shape is the IP, not the math. A napkin version gets you 80% of the value: