Leadership Narrative — MangaAssist

This folder is my leadership story for MangaAssist (Amazon JP Manga storefront chatbot), pitched at the Applied ML Solutions Architect / ML Product Engineer level. It explains how I navigated a project past the 80–90% ML-failure rate, what the data science degree and 2 years at Amazon contributed to the decisions, and how I scaled that judgment across the team.

Read in this order

LEADERSHIP-NARRATIVE.md — the master document. Phase-based journey through the 5 phases of an ML project, with a scorecard mapping every industry failure mode to my mitigation. ~10-minute read.
stories/ — deep-dive STAR stories that elaborate on the moments where math, algorithmic depth, and architecture judgment changed the outcome. Each story is self-contained and ~5-minute read.

Stories

01 — The 92% POC trap — how a curated dev-set masked a 21-point gap and how I designed the eval gate that caught it.
02 — The hybrid execution model — why I refused "use an LLM for everything" and what the cost/latency math actually said.
03 — Debugging the embedding plateau — three weeks of stalled fine-tuning, fixed in four days by reading the loss geometry instead of throwing data at it.
04 — Calibrating the hallucination threshold — turning 0.5/0.7 from defaults into ROC-calibrated decisions.
05 — The EWC retraining save — stopping a 4% regression-per-cycle bleed with a Fisher-information penalty.
06 — The five-persona mentoring framework — how I scaled decision quality across the team without being in every meeting.

How to read these

Each story is structured loosely as Situation → Task → Action → Result, but with two extra layers I care about as a lead:

The math/algorithmic depth that unlocked the decision. This is what the DS degree and the senior Amazon mentors taught me to look for. Without it, every one of these stories ends in "we shipped the trendy answer."
The leadership multiplier. What I taught the team while solving the problem, so the same mistake didn't recur — and so my judgment scaled past me.