AWS Builder Center
I Set Out to Catch Cheaters. I Built Something That Refuses to Accuse Anyone.

I Set Out to Catch Cheaters. I Built Something That Refuses to Accuse Anyone.

Hiring Guard - An interview-integrity platform that turns behavioural telemetry into an explainable score and a list of flags, never a verdict. Show the evidence. Leave the decision to a person.

I am not the smartest person in my batch. What I have is grind.
I sat interviews with nothing open but the editor. No second monitor, no phone below the frame, no friend on a call listening to the question and typing the answer somewhere I could read it. And I watched people who did all of that get the offer before I did.
That is the fire this started with. It was not curiosity. It was not a resume line. I wanted to build a machine that could look at an interview and say: this one is lying.
Four days. Three of us. A hackathon clock and a problem I had been angry about for a year.
Here is what happened instead.



Every product in this space ends the same way

We looked at what already exists. Every proctoring tool we found ends in a verdict. A percentage. A red banner. A pass/fail. "73% likely cheating."
That number is not evidence. It is an accusation with a decimal point on it, and the decimal point is what makes it dangerous — it sounds measured, so nobody opens the session to check.
Think about who that number actually lands on.
The person running an overlay reads the score, understands the game, and adjusts. They learn what the system watches and they stop doing it in front of the camera. A verdict is feedback to them.
The person it destroys is the one who does not know the game exists. Whose room is loud because they share it with four people. Whose laptop drops frames because it is seven years old. Who looks away when they think, because that is what thinking looks like on a lot of faces, including mine.
I had spent a year being angry at a system that could not tell me apart from a cheater. My very first instinct was to build one that could not tell a cheater apart from bad Wi-Fi.
That is the part I want to be honest about, because it is the part everybody skips. The bias was not in the data. It was in me, on day zero, in the shape of the output I assumed I was building.



So we wrote the rule before we wrote the product

The rule: the system produces an integrity score and a list of flags, and every flag lists the observations behind it. It decides nothing.
A sentence like that lasts about three days. Day one everyone agrees. Day three someone is tired, a feature needs a boolean, and  isSuspicious  enters the schema because it makes a component simpler. Nobody is being dishonest. The constraint just has nothing holding it up.
So we made it seven tests, written before the features they constrain:
def test_rule1_no_verdict_field() -> None: ...
def test_rule2_absence_is_never_evidence() -> None: ...
def test_rule3_flags_reconstruct_their_score_delta() -> None: ...
def test_rule4_narratives_never_use_banned_words() -> None: ...
def test_rule5_no_text_enters_the_engine() -> None: ...
def test_rule6_output_always_carries_its_caveats() -> None: ...
def test_rule7_dismissal_only_ever_raises_the_score() -> None: ...
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article