SidharthSundaram

Before I trusted the grader, I made it show its work on every verdict. It got all 12 right — but two of them? Wrong reasoning, right answer. That's why the eval came first. And it paid off: the eval caught three things a pass/fail scorecard would've hidden. The system was stating things it couldn't verify as fact. It was accepting what users told it without checking against real data. And when it refused a question, it couldn't explain why. Those three gaps became the v2 spec. A clean 12-of-12 would've taught me less than the 2-of-12 that cracked these open.

I picked precision before I had any data. The reasoning was simple: a controller who gets five false alarms stops opening the tool. Same way a developer mutes a code reviewer that flags style on every PR. Miss a real commitment? It shows up later when the invoice lands. Cry wolf? You lose the relationship. Then I ran it. Four false positives out of 128 cases. All four were the same thing: budgets sitting under plan, where the agent saw an email-only commitment it didn't expect and flagged anyway. Budgets that were fine. The exact failure mode I'd called before the run.

Built V1 as a confidence coach — silence detection, Whisper transcription, real-time nudges. Tested it with would-be creators. Reactions were... fine. Not excited, not angry. Just flat. That's harder to read than failure. When something breaks, you know what's wrong. When it's lukewarm, the instinct is to keep iterating on the solution. I almost did. Went back and talked to more people, and the pattern was different from what I expected: they weren't freezing because they lacked confidence. They froze because they hadn't figured out what they wanted to say. The bottleneck was 30 seconds before the camera turned on. Rebuilt the whole thing around that.

PostHog was in the first deploy. Before a single user touched it. Not because someone asked "how's it doing" three months in — because if you don't instrument from day one, you're guessing. Page views tell you traffic. Instrumented events tell you whether someone actually finished a course plan, came back, or shared it. That's how I know it's 500+ users and 85% complete the full flow — not a vanity number, not a guess. Zero marketing spend means every one of those users found it on their own. I know that because I was watching from the start.

VP Sales wanted Product Marketing. Pushed hard — sales intuition, strong opinion. Went to my manager. They weren't exactly supportive: "find a cheap/easy solution if you're super confident." So no authority to override the VP, no cover from my own boss. Reframed it around the downside: if I'm wrong, we lose 2 weeks of webinar time. If we skip the test, we risk 6 months of build time. Ran both as webinars. Growth PM drew 3x the demand. After that, nobody needed convincing.

Sales kept telling me buyers could see the technology but couldn't connect it to what their end users actually needed. Looked at the product and had a gut read: this was built for one audience and sold to a completely different one. But gut reads don't move stakeholders. Spent 3 days on LinkedIn finding domain experts. 10 calls. Every single one pointed the same direction. The calls weren't the discovery — I already knew. They were the evidence I needed to make the case.