Behind the scenes · built for KKR

How Exa finds the targets and makes the case.

You enter one company. About 90 seconds later you have a deck of the smaller companies worth acquiring around it, ranked and reasoned. Here is the whole journey, in five steps.

One company
Find similar
Filter
Research
Recommend

01

The job: buy small, build big.

KKR picks one solid company in an industry full of small players (the platform), then buys up smaller, similar companies (the add-ons) and merges them into one market leader worth far more than the parts.

Private equity buy-and-build (a 'roll-up')

02

The catch: the best targets hide from databases.

Those smaller companies are private, often family-owned, with barely a website. No database lists them, because databases mostly track companies that raised money or made news. So analysts spend weeks searching by hand.

PitchBook and Crunchbase track funded companies, not the long tail

03

Exa finds them by what they do.

Give Exa one company and it surfaces the whole hidden field by meaning, the off-database names included. It uses findSimilar for acquisition targets, semantic similarity over the live web instead of a curated list. Then a second, complementary move, the Exa Agent, runs multi-step research from the thesis to pull in the freshest, under-the-radar names that even similarity misses, deduped against the set so far.

Exa findSimilar + neural search, then an Exa Agent deep-research pass over the live web

04

Keep only what's real and buyable.

An AI relevance gate removes look-alikes that only share a name, and an independence gate drops companies already owned by a larger parent (a sub-brand of a national chain is not a deal you can do on its own). What survives is real and acquirable, so it then pulls each company's size, location, and ownership signals, plus one market-size number cited to a real source.

Gemini relevance gate drops name-collisions · independence gate drops parent-owned sub-brands · per-company facts + a cited market stat from Exa

05

It writes the recommendation, as a deck.

One analysis turns the set into a partner-ready recommendation: which targets to call first and why, the value, the risks, and the ask, rendered as a fixed-frame slide deck in KKR's brand in about 90 seconds.

Gemini 3.5 Flash · structured deal thesis · KKR deck · shareable + PDF + CSV

What it runs on

Exa

exa-js · findSimilar + neural search + Contents + Exa Agent · cited market sources

Google Gemini

@google/genai · gemini-3.5-flash · seed 7 · structured JSON output

Vercel

Functions on Fluid Compute · NDJSON streaming route · rate limit + budget guard

Next.js · React

16 App Router · React 19 · server-rendered shareable /r/[id]

Google Firestore

firebase-admin · saved decks for shareable links

PostHog

posthog-js · product analytics

Notes for engineers

  • · Discovery runs two complementary Exa calls: findSimilar (off a domain) builds the known universe, and the Exa Agent API (multi-step research off the thesis) adds the emerging long tail, deduped against it.
  • · The independence gate runs twice: a Gemini pass drops known sub-brands up front, then per-company Exa text is re-checked for a parent/owner, dropping any that slipped through before they reach the deck.
  • · The market stat is sector-scoped and Exa-cited (a single company's ARR is rejected) with a confidence label.
  • · Quant is evidence-only and labeled estimated, blank when unknown. No invented revenue, multiples, or market shares.
  • · Conviction is a derived, calibrated band, not a default: scored from the set's opportunity, edge, and evidence, so a thin set reads honestly as Medium/Exploratory instead of a default High.
  • · The deck streams behind a gated build view and reveals on done, so you never see half-built numbers.

Want the engineering deep dive?

The full backend, every Exa and Gemini call in order, the streaming protocol, and the trust layer, is laid out on its own page.

Read the full architecture

Architecture overview · informational. Back to Exa Vantage.