Product One
The Voice Interviewer
This is the one that started everything. Sal came to us first with the front door: the
mentor application itself.
The Problem
GLEAC's application ran close to an hour. More than half the people who started it never
finished. And a growing share of the ones who did were pasting in ChatGPT answers, polished,
generic, and useless for judging whether someone actually knew their subject.
GLEAC's own screening algorithm was catching over half of submissions as inauthentic and
rejecting them on the spot.
"It was a very clear business problem statement that we had to solve. The outcomes if we
solved it well was going to be less drop off and a higher volume of people we could
onboard."
Sal, Founder
The Goal
Sal's target was aggressive: cut onboarding from an hour to under ten minutes, roughly six
times faster. Double completion rates. Get 90% of profiles to at least 80% complete without a
manual follow-up. Her answer to all of it was voice. A live conversation is hard to
fake on the spot, in a way a text box never was.
That single shift, from a form to a conversation, is what took onboarding from 45 minutes
down to under ten.
Every question built from the last answer
Generated live
Follows the last answer
One applicant, speaking
No two interviews alike
A script asks the same thing every time. This one listens first.
Why This Was Hard Core
We started building in November 2024. Nothing about it was routine, and most of what made it
hard had nothing to do with the AI itself.
Personalizing every interview in real time.
Any team can wire up a voice API and call it an interviewer. Ours had to do more. It
asked a different question to every applicant based on their actual background. It
followed up on what they'd just said instead of running down a script. Every question was
generated live, inside a conversation with no retakes. That's the part that turned a
working prototype into something that could actually hold a conversation, and it's the
part that took the most time to get right.
Making the voice feel like a person, not a system.
GLEAC's applicant pool skews Gen X. Most were talking to an AI interviewer for the first
time in their working life. We had to get the pacing and the tone right. Feel robotic or
move too fast, and people just stopped talking. That was a product design problem before
it was ever a technical one.
Shipping through a holiday break, on a tight window.
The build ran six to eight weeks end to end. That window included a full break over
Christmas where neither side was working. Everything downstream had to happen in whatever
time was left, with no room for a second pass at the parts that turned out harder than
expected.
Product Two
The Mentor Evaluation System
Fixing the front door only mattered if GLEAC could actually process what came through it.
That's the second fire.
The Problem
Every mentor application went through three human evaluators, in a batch process that took
three to four weeks. That capped GLEAC's entire pipeline at 25 to 30 new mentors a
month, no matter how many qualified people applied. The work itself was repetitive
and manual: exporting answers, building evaluation sheets, running scoring meetings,
invoicing evaluators by hand.
The Goal
The target was automation, not incremental speedup: cut manual evaluation effort by roughly
95%, and bring the approval timeline down from 30 days to same-day, ideally within hours. A
secondary goal mattered just as much to Sal: consistency. Three human evaluators
scoring the same application differently was its own quiet problem, one an automated
system could remove almost entirely.
Same application, three opinions
Three evaluators
3–4 weeks
A
B
C
A spread nobody could explain
Automated
Same day
One reading, every time
The wait was the visible problem. The disagreement was the quiet one.
These are the goals GLEAC set out to hit with this system. Confirmed, audited results for
this specific product weren't part of what Sal covered in her own account of the engagement.
Why This Was Hard Complex
Evaluation isn't pattern-matching against keywords. It's judgment: does this answer actually
demonstrate the expertise being claimed, or does it just use the right vocabulary. Building a
system that could make that call automatically, at a consistency rate close to what three
trained humans agreed on together, meant the model had to reason about answer
quality, not just check for the presence of relevant terms.
Product Three
The Profile Podcast Pipeline
Not every product on this list came from a business problem. This one came from an instinct
Sal had about how new mentors should feel on day one.
The Idea
The moment a mentor gets accepted, the system pulls their LinkedIn profile and produces a
two-host conversational podcast about their own background, fully automated,
live on their profile before they've done anything else. A Lambda function
fires on acceptance and the whole pipeline runs with no manual step.
Live before day one begins
Shared onward, unprompted
No bottleneck was removed here. It just made a first day feel like something.
When Sal first pitched it, our own read was that it was technically trivial, a thin wrapper
around NotebookLM, nothing groundbreaking. We were wrong about what mattered.
Sal's read on it was the opposite, and she was right. Mentors started sharing their own
podcast across their networks unprompted, something no other onboarding step at GLEAC
does.
Why This Worked Delight
This wasn't a product built to fix a bottleneck. It was built to make a first day feel like
something. That's a different kind of hard to justify technically, and it's exactly the kind
of call that only made sense in hindsight.
