Sabashan Ragavan

The through-line

Ragavan's consistent preoccupation is that large language models have arrived on both sides of the hiring table at once, and that recruiting technology has to make sense of that fact rather than pretend it away. On the candidate side, he argues AI is "a tremendous accelerant to one's productivity" and that applicants should not be penalized for using it, while employers still need a way to tell productive use from outright cheating on screeners [1]. On the employer side, his product thinking runs in the same direction: put the model to work reading the job description, generating the questions, scoring the answers and adapting live, so that a structured interview costs a recruiter a few minutes of setup [2].

On assessing candidates who use AI

The core question he sees reshaping engineering hiring is "identifying how candidates can actually use AI when they're building" [1]. His position is that the tooling has changed the job, so it should change the assessment: "I don't think candidates should be, you know, potentially penalized if they're using an AI tool to help them code more effectively," and "there needs to be ways to assess the tenant's ability to leverage these different tools" [1]. He frames this as a straightforward consequence of the environment rather than a concession, noting "we're in a different world now that, uh, you know, large language models" [1].

The counterweight is detection. Alongside measuring skilled AI use, he expects "a rise of employer, you know, using solutions to catch those candidates" whose output is simply "coming up as a response of using an LLM versus using it to, you know, actually be better at their jobs" [1]. The distinction he insists on is between the candidate whose work improves because of the tools and the candidate whose screener answer is nothing but the tool's output, and he treats separating those two as the central unmet need in assessment technology [1].

On generating interviews from the job posting

In practice, his answer to interview design is to remove it from the recruiter's hands as a starting point. He shows the workflow as pasting a job title and the full job description straight from a public job board into HeyMilo, selecting an interview type, with "technical" described as the most common for the agents, and letting the system read the description "to come up with the initial set of baseline questions to ask candidates as well as an evaluation criteria for each of these questions" [2]. The generated set for a senior program manager role covers familiarity with labor employment laws, use of MS Office applications, handling a challenging key account issue, profit and loss management, and operations experience in the workforce solutions space [2].

Scoring is built in at the same moment as the question, not bolted on afterwards. Each question arrives with a visible rubric showing "what a score of one will look like versus a score of five" [2]. The generated set is treated as a draft rather than a verdict: every question can be modified, new questions added from scratch, and any question deleted [2]. Agent behaviour is similarly tunable through a free-text customization field, with instructions like pausing after an initial question or speaking very professionally offered as examples [2].

On adaptive interviews and frictionless distribution

He is explicit that the agent does not simply read out a fixed list. Once candidates begin interviewing, "hey Milo will still adapt to the candidate's responses and ask follow-up questions" [2]. Distribution is deliberately minimal: setup produces a single link that serves every applicant to a posting, which can be embedded directly in the job posting, emailed to candidates who apply, or integrated with existing ATS systems, whichever the employer prefers [2].

Takeaways

• Candidates who use AI to code more effectively should not be penalized for it; assessments need to measure the ability to leverage those tools [1].
• The unsolved problem is separating genuine AI-assisted work from screener answers that are purely LLM output, and employers will increasingly buy solutions to catch the latter [1].
• HeyMilo generates baseline interview questions and per-question evaluation criteria directly from a pasted job title and description, with technical the most common interview type [2].
• Each generated question ships with an explicit rubric spelling out a score of one versus a score of five, and every question can be edited, added to or deleted [2].
• Agent behaviour is customizable through free-text instructions such as pausing after an initial question or speaking professionally [2].
• The agent adapts to candidate responses and asks follow-up questions rather than running a fixed script [2].
• One interview link covers all applicants to a posting and can sit in the job ad, in email, or in an existing ATS integration [2].