Technical Program Manager

Opening Statement (~60 sec)

I owned the cross-functional program to fix a broken ad-deduplication system inside Amazon's Prime Video and Freevee AdTech stack — a system I had inherited as 'already solved' but that was actually failing in production. The existing Creative ID matching system was catching only ~2% of true ad repetition; viewers were seeing the same ad up to five times an hour, CSAT had dropped 8% over six months, and the issue had escalated to board level. Rather than keep presenting the failures as edge cases, I owned the diagnosis publicly, built the cross-functional case for a full AI/ML rebuild, and ran the three-team delivery — ML Science, AdTech Infrastructure, and Ad Ops — through to a validated global launch. The result: true repetition detection went from 2% to 40%, global detection accuracy reached 98.5%, the fix launched across 17 markets, and the program recovered $87M+ in annualized revenue — while the phased, trust-building rollout I designed converted the most skeptical engineering stakeholders into co-owners.

Situation

1. As the TPM for ad quality on Amazon's Prime Video and Freevee AdTech platform, I inherited a program the org considered closed: a Creative ID-based deduplication system built and shipped to stop viewers from seeing the same ad repeatedly within a session. 2. On paper it looked like it was doing its job — it matched ads by exact file/asset ID. In practice it was catching only ~2% of actual repetition, because it couldn't recognize a re-uploaded creative under a new Creative ID or link two file variants of the same commercial (for example a 15-second and a 30-second cut) as the same ad. 3. Viewers were seeing the same commercial up to five times in a single hour. Customer satisfaction scores had fallen 8% over six months, repetitive ads became a top complaint in app-store reviews, and the issue escalated to board level. 4. This was a program under my ownership that was quietly failing — I had been reporting the gaps in steering reviews as 'coverage edge cases we're tuning' rather than naming the root cause: the system was solving the wrong problem.

Task

1. My task was twofold: run the technical diagnosis to confirm whether the existing system itself — not advertiser behavior — was the root cause, and then decide, and own, what to do about it. 2. A VP gave me direct feedback before the next steering committee: I had the evidence the current approach wasn't working, but kept framing it as an optimization problem to protect stakeholder comfort. My task shifted from managing the existing program to calling the diagnosis clearly and leading the org to a new solution. 3. That meant owning a full cross-functional rebuild — from root-cause validation, through a technical case for an AI/ML-based system, through delivery sequencing across three previously independent teams — without an existing playbook or mandate to redesign a system the org had already accepted as solved.

Action

1. 1.Ran an independent root-cause analysis of the Creative ID system — outside any assigned mandate — and documented three structural failure modes: identical ads re-uploaded under new Creative IDs during campaign refreshes, same-commercial cuts (15s/30s) never linked because their file metadata differed, and a measured ~2% true-repetition catch rate against a later-confirmed ~40% actual rate.
2. 2.Reframed the problem publicly at the next steering committee instead of continuing to call it a tuning gap: stated plainly that the system matched file assets, not what viewers actually perceive as "the same ad," and that the fix required a semantic-similarity approach — then walked the engineering leads who had built the original system through the failure data so the conversation stayed evidence-driven rather than defensive.
3. 3.Structured the technical redesign as a cross-team co-design, not a top-down spec: ran two weeks of working sessions with ML Science, AdTech Infrastructure, and Ad Ops to jointly define a multimodal pipeline — speech-to-text, logo/object recognition, and an LLM layer extracting brand, product, and tone into a vectorized "ad signature" scored by cosine similarity.
4. 4.Personally ran an Architecture Decision Record (ADR) session with the infrastructure team to formally capture the trade-offs they hadn't been consulted on before the pivot — converting a team that had built the original system, and had reason to be defensive about it, into a co-owner of the replacement.
5. 5.Sequenced delivery to rebuild trust before asking for it: Phase 1 shipped measurement-only — profiling every ad, computing similarity scores, and publishing them to a dashboard with no filtering — so skeptical stakeholders could validate the new system's findings against the old system's blind spots before any enforcement went live.
6. 6.Owned the global rollout plan across 17 markets once Phase 1 validated the approach, sequencing regions to manage localization and compliance risk rather than launching every market simultaneously.

Result

1. True ad-repetition detection rose from ~2% to ~40% — the Phase 1 measurement-only data made the business case undeniable before any filtering shipped. 2. The rebuilt system reached 98.5% global detection accuracy and launched across 17 markets. 3. The program recovered $87M+ in annualized revenue by closing the gap between reported and actual ad repetition. 4. The infrastructure team that had built the original system, and had the most reason to be defensive about replacing it, became active co-owners — two engineers who had been passive contributors began surfacing model improvements unprompted, including a self-verification check that took missing brand labels from 8% to 0%. 5. The ADR process I introduced to capture the infrastructure team's trade-offs was adopted by two other programs in the org as a standard practice for mid-program architecture pivots.

Closing Statement (~60 sec)

What made this a program-management failure to own, not just a technical one, was that I had the data before I had the resolve to call it — I'd been softening a structural finding into a tuning narrative to avoid reopening a decision the org considered closed. The fix wasn't just the AI/ML rebuild; it was naming the failure plainly, in public, before I had a fully-built replacement, and then designing the delivery — the co-design sessions, the ADR, the measurement-only Phase 1 — so the teams who'd built the original system had a way to move from defending it to improving on it. The pattern I carry forward: when a system I own is failing, I report the root cause, not the most comfortable interim explanation — and I sequence the fix so trust is earned with evidence before I ask anyone to accept a bigger change.