Quality: For the example of "session summarizer," I would like to optimize the quality in terms of usefulness and reliability, as the output is variable and there is no single correct answer. I would start by defining what a baseline good-quality answer is by checking whether the summary helps the users without rewatching or asking the same question again with minimal changes to the prompt. Avoiding any hallucinations and providing a clear structure and grounded summaries based on sources. Use guardrails by setting length limits and omitting when unsure. Leverage feedback loops by checking whether the user found the output useful or not. Initially, apply human review to set good baseline standards before scaling it Success: Success cannot necessarily be defined by measuring the accuracy of the output. We could use some metrics, like how useful the user found the output by measuring the user engagement with the copy, save, or share CTA. We could also measure whether the user takes the first response or continues with more prompts to make the output more efficient to their liking. Additionally, calculate the repeat usage across the sessions to understand the adoption. Ensure the quality of the response with Helpful/ Not helpful response from users or monitoring bad summary rates over time. Finally, monitor the users' use of the system even though they are not satisfied with the initial output. North Star: Number of sessions where the summary becomes the user’s primary takeaway per Total sessions

PROBLEM In my previous organization, teams would often make important decisions repeatedly but didn’t learn which ones were good or bad, and importantly, the reason behind them (why?). Q1 – WHAT PROBLEM The decisions live across different scattered documents, emails, meetings, and Slack threads. Outcomes happen weeks or months later, disconnected from the original context. As a result: - The same bad decisions are repeated - Good Judgment isn’t recognized - Retrospectives focus on outcomes, not decision quality under uncertainty Teams usually optimize for speed and alignment without focusing on the learning aspect. Q2 – SHOULD WE USE AI AI is essential in this case because decision quality is contextual and retrospective. Simple documentation or checklists can capture what was decided but rarely capture the WHY, constraints, or how decision-making took place over the course of meetings. AI is needed to connect decision context to outcomes over time and across teams. Q3 – DO WE HAVE THE DATA Yes, but it’s fragmented and exists across: - Documentation (Decision docs, PRDs, RFCs) - Emails, meeting notes, and async threads - Metrics and business outcomes - Reversals, rollbacks, or follow-on decisions The data exists; it’s just never linked. Q4 – TRANSLATE TO A MODEL The model reconstructs decision narratives: - What assumptions were made - What signals were considered or ignored - How outcomes compare to expectations Outputs will be the decision quality patterns (e.g., “decisions made with X signal tend to succeed”)—not judgments of individuals. Q5 – USER EXPERIENCE The management and PMs see a decision timeline: context → options → choice → outcome → learning. This tool answers, “What should we repeat? What should we avoid next time?” The focus is entirely on learning instead of individual decision-making Q6 – LAUNCH & QUALITY Good enough when teams reference it during planning and retros. Early success = fewer repeated mistakes and clearer rationale in future decisions.

Late to the party but here's my opinion PRODUCT NAME: Spotify recommendations WHAT IT DOES: Suggests songs that are suited to the customer's taste based on their song preferences. Additionally, source songs that are similar/complementary based on other users from its dB WHAT CAN'T BE FULLY SPEC'D: The logic is not easily deterministic and is very complex. It is very difficult to be certain about what the customer actually likes or prefers. The song choices are not a binary choice; multiple songs could complement the customer's liking, but suggesting the most appropriate one is where the optimum value lies WHY AI (NOT SOMETHING SIMPLER): For the abovementioned reasons, something simpler would limit the song choices and prevent exploring newer datasets across broader data points. Handling data at scale and sharing customized recommendations is where AI is preferred HOW IT HANDLES BEING WRONG: The customer is given a choice to dislike the song and notify the model, which takes it into account and will not necessarily repeat that song as a recommendation. ONE THING THE TRADITIONAL PM WOULD MISS: The agility of the AI model to ramp up on newer trends and incorporate the newer customer behavior changes is something that would be missing from the traditional PM approach. The challenge in the traditional approach is there is no fixed outcome, so you cannot define an exact definition of a perfect recommendation.

any recorded sessions for people who are in a different time zone?

@Sannidhi Pushpey thank you. I'll check out the recording of this session

The biggest challenge has been getting past the ATS system. I believe I have conveyed the impact in STAR format, and all the necessary details are present on the resume. However, I'm not sure what I'm missing to get past the ATS. Would love to access the resume booster and if anyone could help with a quick resume review,

@Sannidhi Pushpey Thank you for the support. I have messaged you and will keep you updated with my progress