Umamy Replaced Exa with Seltz for Deeper Candidate Research

At a glance

• Customer: Umamy, Paris based agentic recruiting platform

• Use cases: Scorecard enrichment and lookalike candidate search

• Previous approach: Exa-powered web search within candidate scorecards

• Why Seltz: Full candidate profiles and structured company data, deep enough to resolve the scorecard criteria that a profile alone leaves open

• Result: Both workflows in production two weeks after the pilot began, built jointly with the Seltz team

Umamy's end-to-end recruiting platform

Recruiting teams need more than a list of people who match a job title.

For Umamy, the work is helping startups assess whether a candidate meets the specific criteria behind a role:

• experience at a VC-backed company

• a particular technical background

• education requirements

• context about a lesser-known employer

Some of that information is available in a candidate profile. Some of it requires research beyond the profile. That is where Seltz now fits into Umamy's platform.

Umamy uses Seltz for scorecard enrichment when existing candidate data is insufficient, and for lookalike candidate searches based on a reference profile.

The challenge: candidate profiles do not answer every important question

Umamy's sourcing and filtering workflow already worked without web search.

The gap was in scorecard enrichment: validating criteria that may not appear clearly in LinkedIn or other profile data.

The team identified several examples:

• Whether a candidate had worked at a VC-backed business

• Whether a university met a ranking threshold, such as top 50 in U.S. News & World Report

• How to classify a lesser-known company, such as a boutique consultancy

• Technical-stack information for engineers whose LinkedIn profiles were sparse

Umamy had been using Exa for this research. The limitation was depth. Confirming that a candidate's employer exists is a different problem from establishing whether that employer is venture backed, how large it is, or what it actually sells. Snippet-level results identify the entity. They rarely settle the criterion.

"We did not need web search for every candidate or every scorecard item. We needed it when the information already in the profile was not enough to make a confident call."

Jonathan Bouaziz, Umamy CPO

A more targeted approach to scorecard enrichment

Umamy redesigned the flow so that an LLM first evaluates a scorecard using existing data. It identifies gaps, calls research only for unanswered items, then re-scores the candidate with the additional information.

For scorecard enrichment, Umamy also moved from triggering research per criterion to triggering it once per person when needed.

That distinction matters for answer quality. Routine criteria, such as whether a candidate has TypeScript experience, can usually be resolved from profile data. Research is reserved for the criteria that require external context, which are also the criteria where retrieval depth determines whether the scorecard gets a real answer or a guess.

Why Umamy evaluated Seltz

Umamy evaluated Seltz alongside Exa using the same prompts in a lightweight benchmark, comparing outputs on representative queries drawn from live scorecards.

Two capabilities were decisive:

Full candidate profiles rather than snippets. Seltz returns the complete professional profiles: full work history with tenure dates, projects, and skills. For a scorecard question about a sparse engineering profile, or about whether two candidates overlapped at the same employer, the answer lives in parts of the profile a snippet truncates.

Structured company data for employer criteria. A large share of scorecard criteria are questions about the employer rather than the person: funding history, headcount, whether the business is privately held, what the company actually does. Seltz's company index returns that as structured data, alongside executives, job posting titles, tech stack, and where a company hires from and loses people to.

Companies change quickly. Rounds close, headcount shifts, hiring plans change.

Umamy caches results after the first lookup and invalidates them on a schedule for exactly that reason: a criterion about employer size or funding stage is only as good as the last time the underlying record was refreshed.

Seltz re-crawls its company index on a fixed cadence, so a re-enrichment against the same company returns the current state rather than a repeat of the first answer.

Two weeks from pilot to production

In early July, Umamy outlined its scorecard-enrichment use case and proposed a two-week pilot to test representative queries.

Over the following weeks, our teams worked together to:

1. Define the scorecard-enrichment and lookalike-search use cases.

2. Share representative queries and test Seltz against Exa on the same prompts.

3. Refine the architecture so research runs only where profile data leaves a meaningful gap.

4. Move both workflows into production.

This was not a documentation handoff. The Seltz team worked directly with Umamy's engineers throughout: opening up additional scopes as new questions surfaced, writing cross-prompts for company search, supporting the benchmark, and offering to help build an evaluation layer for the test pipeline. When results did not come back as expected, we debugged the query rather than pointing at the docs.

Umamy and Seltz also worked through a blue-green-style testing plan, routing comparable queries to each provider and comparing results before broadening use.

How Umamy uses Seltz today

1. Scorecard enrichment

Umamy calls Seltz to enrich scorecards when existing profile data does not answer a required criterion. That includes determining whether an employer meets a defined company criterion, such as being small or privately held. Results are cached after the first lookup and refreshed after a few months.

This gives recruiters real evidence on the criteria where a candidate profile alone cannot establish the answer.

2. Lookalike candidate search

Recruiters can provide a reference profile, for example an AWS engineer, and use Seltz to validate the profile and build a hunt around it.

The broader lookalike workflow combines a seed profile, research, and structured query translation to surface similar candidates. Because Seltz returns complete profiles, the workflow can reason over actual tenure and role history rather than inferring from a fragment.

"Seltz gives us a way to go deeper when the profile alone does not tell us enough, without making web research the default for every decision."

Connor Blair, CTO, Umamy

The result: deeper answers on the criteria that matter

Umamy now resolves scorecard criteria that its previous setup left open, and it does so on a research flow that fires only where profile data falls short.

Two changes produced that:

1. Research depth. Full candidate profiles and structured company data answer employer and tenure questions that snippet results identify but do not settle.

2. A more disciplined architecture. Research runs when scorecard evaluation finds a gap, which keeps every call attached to a question that matters.

For a recruiting product, this matters because candidate fit depends on details that are often incomplete or absent in a profile. Umamy can add evidence where it improves a scorecard, while keeping profile-based evaluation as the default.

What's next: company context at onboarding

The next workflow Umamy is building moves research earlier, into onboarding.

The goal is for a new account to arrive already understood. From a work email at signup, Umamy wants to establish what the company does, whether it has raised, how many people work there, what it is currently hiring for, and who inside the account is likely to be involved in recruiting. That context becomes organization memory, so a new user reaches a working search before entering a credit card.

Onboarding raises the accuracy bar. A scorecard criterion is evaluated by a recruiter assessing a stranger. An onboarding summary is shown back to the person who knows the company better than anyone, and a month-old view of a business that raised last week is the kind of error that gets noticed immediately. Funding events, headcount, and open roles are the fastest-moving fields in a company record, and they are the three the onboarding flow leans on most.

That workflow depends on getting current company information quickly, at signup speed. It maps directly onto the Seltz company index, which returns funding history, headcount, executives, job posting titles, tech stack, and talent flows in and out of a business as structured data, alongside the company's own website and blog content.

We're committed to helping Umamy implement features that give founders and recruiters the power to assess candidate fit with richer evidence.

Interested in trying yourself? Grab an API Key and take a look at the docs.