01
2020 — PRESENT
Tech lead for Fire TV Catalog Enrichment, previously client-side tech lead on Fire TV Partner Integrations. Promoted to Sr. Software Engineer (Q4 2022).
#PROJECTSUMMARY
01Drove the architectural decision to publish to a separate output stream rather than extending the core catalog, unlocking a 10x data-volume ceiling and a faster schema-evolution path for novel data types.
02Formalized a platform model where consuming teams own domain-specific extraction logic on Catalog Engineering's governed rails (lifecycle, linking, publication).
03Aligned the Fire TV Search team to build its Prime Video scene-data curation app on the new platform, setting precedent for centralized data governance across downstream teams.
04Delivered to production solo in three months against a documented 4.5-month estimate. Enabled by a production-like integration test harness and agentic engineering.
01Built an LLM genre classifier embedded in the catalog's synchronous stream-processing pipeline at ~$0.001 per entity; engineered an eligibility filter scoping inference to the ~5% of incoming updates most likely to surface on customer screens, keeping long-tail updates off the latency-sensitive critical path.
02Invented CMAR (Catalog Metadata Acceptance Rate): a measurement methodology using an LLM judge calibrated against PM ratings (>90% agreement on a 10K-entity sample) to validate enrichment quality. Classifier hit 95% CMAR vs. 82% baseline; the metric earned PM confidence and unlocked downstream customer-team adoption of LLM-classified genre values.
03Tech lead and design approver on subsequent enrichment features (location extraction from linear-station metadata, image metadata extraction) shipped by mid-level engineers under reduced headcount; core contributor to the catalog org's shared Bedrock integration library (multi-account rate cycling, prompt construction, JSON parsing, retry, metrics).
01Drove the decision to build the on-device middleware in Rust rather than piggyback on existing Android services. Shipped to 100M+ devices at <5MB memory footprint, proving out the native-service pattern subsequently adopted across Fire TV.
02Drove the decision to build a custom Amazon SDK rather than adopt Android's standard Watch Next API, arguing it would compromise integration quality and partner ergonomics. Owned the entire partner-facing surface end-to-end (SDK, integration docs, reference sample app); the resulting API surface became a partner-cited deal-closing factor across Disney+, HBO Max, Hulu, Starz, and MGM+.
03Built and operated the cloud ingestion backend for sensitive customer and partner data, handling up to ~200K TPS observed peak in production.
04Owned the privacy/consent surface end-to-end. Opt-in capture, propagation through the pipeline, and statistically rigorous experimentation that drove opt-in from 30% → 75% over a few months.
01Scaled platform capabilities across 40+ worldwide Live TV partner launches (e.g., DirecTV), supporting a 2x increase in monthly active users of the Fire TV Live TV experience from 2020 to 2023.
02Built anomaly-detection alarming that surfaced customer-impacting issues in real time.
03Expanded the partner developer ecosystem: major improvements to the reference app and a step-by-step integration guide used by launching partners.
01Authored and maintain catalog-engineer: a shared agent spec (context + MCP servers + skills) adopted 100% across a 3-team, ~40-person org, cited by VP as a best-practice example, and replicated by sister Fire TV orgs.
02Built and operate an internal human-gated autonomous engineering pipeline that ships production code through Amazon's standard peer code review. Q1 2026 alone shipped ~70% as many merged PRs as all of 2025 (~3x output multiplier). Shared the system in an org-wide talk to 100+ engineers, driving 10+ direct adopters and two VP follow-ups.
03Trusted by Fire TV VP leadership to help provision a new AI Engineering velocity team; subsequently drove Catalog org adoption of a code package context graph system, landing 100 core packages in 30 days, setting the implementation standard, and operating a live adoption-tracking dashboard powered by the context graph traversal MCP.