AI evaluation criteria – Tezi

In the second step of calibration, set up AI evaluation criteria—complex, natural language criteria that Max assesses using higher-level reasoning on a resume. AI evaluation complements filters to fully capture the nuances of your target candidate profile.

What kinds of criteria should I use AI evaluation for?

Use AI evaluation for criteria that:

Require analysis, reasoning, or judgment to assess. Simple, factual checks belong in filters.

Example 1: “Current individual contributor” should be an AI criterion because it requires understanding what an individual contributor is and determining whether a candidate is an IC. In contrast, “current job title is [list of IC job titles]” does not require reasoning and should be applied as a filter.

Example 2: “Has experience selling software to national retailers” should be an AI criterion because it requires understanding what qualifies as “software” and “national retailer”, then recognizing that experience in a candidate. In contrast, “has worked at [list of companies that you know sell to national retailers]” can be applied as a filter.

See below for more examples.

Cannot be configured in filtering. Filters run on structured fields in Tezi’s database, so they’re faster and more consistent, and should be used wherever possible. However, because they’re limited to fields Tezi has pre-structured, some requirements can’t be represented with filters. In these cases, use AI evaluation instead.

Span multiple types of filters. While you can create custom criteria that combine logic across multiple filters, advanced combinations are often easier to implement as AI evaluation criteria.

What information is available for AI evaluation?

For AI evaluation, Max knows:

Information you can find on a candidate’s profile – their job titles, experience summaries, skillsets, education, and more.

For sourcing, Max has access to publicly available profile information aggregated from a wide range of online data sources, including LinkedIn, GitHub, and social media profiles.

For inbound screening, Max uses the candidate’s resume or, if they did not submit a resume, their LinkedIn profile.

Information about the companies the candidate has worked at – what they focus on, funding information, company size, etc.

General background information about common job roles and experiences, skillsets, and more.

Think of Max as having a huge amount of information about the world and about the candidate, but not yet as a trained, expert recruiter in a given field. For example, Max can recognize a front-end engineer from their job titles, skills, and projects, but it does not know that a candidate working on a specific team at a specific company or describing their work in a more ownership-oriented way correlates with being a high-quality candidate.

💡

Many candidates may not provide extensive information on their profiles. Max won’t be able to guess about someone’s experiences or projects if the information is not present (e.g., a particular tech stack). Based on your understanding of target candidates, consider the types and depth of information that they would likely share on their public profiles, and focus your AI criteria on those areas.

Best practices

1️⃣ Be objective and specific

Phrases like “strong promotion trajectory” or “has driven significant impact” can have varying interpretations among humans and with Max. Ensure Max’s interpretation is aligned with yours by being objective and specific, such as “promoted to senior SWE, staff SWE, tech lead, or engineering manager within the first 5 years of their career”.

💡

Specificity is important, but only be as specific as your real needs are. Otherwise, you may artificially constrain your candidate pool.

2️⃣ Understand impact on pool size

After you retrieve matching profiles, our system provides an estimate of all matching profiles. Compare this estimate to the number of matches for the last filter. The difference is the impact from AI evaluation.

If AI evaluation drastically shrinks the pool, consider whether all criteria are essential or some can be removed. You can also test AI criteria individually to see which ones reduce the pool the most.

3️⃣ Prior to AI evaluation, use basic criteria to filter the candidate pool down to <5,000 candidates

Max does an in-depth evaluation of each candidate profile that meets filters in order to assess whether they match AI criteria. If too many candidates pass the filters, Max can struggle to analyze all profiles, and AI evaluation may fail.

Prior to AI evaluation, filter down to a pool of <5,000 candidates to ensure Max is able to complete its AI reviews.

4️⃣ Consider the interaction between filters and AI evaluation criteria

If an AI evaluation criteria can be represented through filters, use filters instead to produce faster, more consistent results.

Avoid incompatible criteria between filters and AI evaluation. For example, if you want to use AI to identify top-performing sales representatives, make sure that your filters first select for job titles that top sales representatives would likely have.

Examples

Experience in a specific domain

Bad examples	Good examples
Candidate is a backend engineer	Candidate is a backend engineer focused on API services and backend projects for web apps
	Candidate is a backend Python engineer focused on technologies like (but not exclusive to) Django, FastAPI, …
	Candidate is a distributed systems backend engineer focused on building backend infrastructure in C++, Go, Rust, or similar
Candidate is a growth product manager	Candidate is a growth product manager working on retention, and has demonstrated experience with A/B testing
	Candidate has experience as a growth product manager focused on new user acquisition and activation
Candidate has worked in insurance sales	Candidate has experience selling personal lines of insurance (auto, home, life, etc.) online
	Candidate is a sales agent focused on corporate liability insurance, and is licensed in California
Candidate has worked in healthcare	Candidate has experience working with consumer fitness wearables
	Candidate has experience in pharmaceutical research and oncology
	Candidate has experience coordinating clinical trials for vaccine studies
Candidate has 4 years of experience as a head of marketing at a series seed to C B2B startup	Filters • 4 years of experience as a Head of Marketing, Chief Marketing Officer, or Director of Marketing • Has worked at a Seed to Series C Startup These filters help narrow down to the right target pool. AI Reasoning • Has 4 years of experience as a head of marketing at a series seed to C B2B startup Use the same AI criterion as before but achieve better results by applying relevant filters first.

Without specificity, Max will be confused about what kind of candidate you're looking for and will return candidates that aren't eligible for your role. In the first four bad examples above, Max won’t know what type of backend engineer (security, distributed systems, Python services, etc.), growth PM (Acquisition, Retention, Monetization, etc.), insurance sales (consumer, retail, online, etc.), or healthcare experience (consumer fitness wearables, oncology, clinical trials, etc.) you're looking for.

If you require expertise at that level of specificity, provide it to Max. Conversely, if you don’t require that level of specificity, do not include it in your criteria to avoid artificially constraining your pool.

In the last example, when your AI evaluation criteria are highly specific, try to set filters that narrow down on the right target pool. This will speed up AI evaluation and produce more consistent results.

Candidate quality

Bad examples	Good examples
Candidate is a rockstar engineer	Candidate demonstrates shipping multiple projects to production.
	Candidate shows a metric (revenue, growth, performance improvement) outcome from their work.
	Candidate has shipped lots of projects quickly and lists them on their resume
Candidate shows great ability to sell	Candidate shows track record of success (i.e., look for things like hitting quota, president's club, etc.)
	Candidate has closed deals worth over $250,000
	Candidate owned their full sales cycle from lead generation to closing deals
Candidate has been published	Candidate has published papers on oncology in a peer reviewed journal
	Candidate has written articles published in well-known tech magazines

For subjective criteria like "rockstar" or “great ability to sell”, Max won't know what you mean by “rockstar” or “great”, so being more objective and specific is helpful.

For broad areas like "has been published", Max will look for any publications they’ve listed, e.g., even a kids book they self-published. Try to be specific on what type of publishing you're looking for.

Candidate career history

Bad examples	Good examples
Candidate has a successful career	Candidate has listed large, impactful projects they’ve shipped
	Candidate has moved from a junior role to a senior role within the first 2 years of their career
	Candidate has worked as a director level or above at a well known tech company
Candidate hasn’t worked at an (agency/startup/company)	Hasn’t worked at a design agency in the past 6 years
	Hasn’t worked at pre IPO startups for the past 5 years
	Candidate hasn’t worked at Google since 2010

Targeting broad terms like "hasn't worked at an agency" will exclude candidates who currently work at agencies and who worked at one 15+ years ago. If you’re open to candidates without recent experience at an agency, try being more specific.

Detailed candidate experience

Bad examples	Good examples
Has experience mentoring others	Has led teams as a people manager or senior tech lead
	Has been involved building or shaping internship programs
	Has been part of speaking panels or given talks on brand marketing

In this example, if you need a specific type of mentorship experience, spell it out for Max. Otherwise, Max may surface candidates whose mentorship experience doesn’t meet your needs.

Note that you may struggle to find matching candidates for detailed experience criteria, depending on the level of detail most people provide in their profiles. In those cases, consider evaluating for the experience in the chat screen or an interview.