Felipe Duenas

June 2026 — Present

Now

Software Engineer Intern

Amazon Web Services · Data Platform · Seattle, WA

• The org's operational knowledge lived in 3,000+ Tableau dashboards that no software could read. I built the pipeline that turned them into something an AI agent can answer from for our internal supply chain automation platform (serving 5,000+ users). 85% more of the org's data became reachable, and the accuracy gain traced to the data, not the prompt.
• Then the other direction: tooling that builds a dashboard from a prompt, so a planner gets one in three minutes as a self-serve request instead of waiting four days on a BI engineer's build cycle. The page is constructed in code rather than typed by the model, which took output cost from 10k tokens to a few hundred and means every cell traces to a real row.

Jan 2026 — May 2026

Backend & Data Engineer Intern

UCLA Trustworthy AI Lab · Agentic Data Science · Los Angeles, CA

• Built the tool-calling layer for the Lab's agentic data science platform, which turns a plain-language ask and a raw file into runnable Python, charts, and an exportable table. Achieved a 65% execution accuracy on BIRD, 8 points above GPT-4o.
• The agent worked due to the data architecture upstream, with an ingestion pipeline that infers schema and column types from any CSV, JSON, or Parquet, and a Postgres event log recording every response and failure. Querying that log and acting on the findings is what cut the failure rate from 18% to 5%.

Jun 2025 — Sep 2025

Business Intelligence Engineer Intern

Amazon Web Services · Arlington, VA

• Planners at government sites were coordinating builds around power numbers padded by buffers on top of buffers. An XGBoost forecast, blended with a Random Forest and fed peak events as their own signal, cut error from a 6.3% baseline to 1.9% MAPE: tight enough to recalculate the planned values and release 21% more usable capacity inside the same power envelope. This led to roughly $400K of rack spend avoided across 33 AZs.
• Underneath it, the data work: Glue loading 500M+ daily telemetry events into Redshift, dbt modeling them into incremental marts, one capacity-planning layer in place of the seven sources my team was reading by hand, and a Bedrock text-to-SQL interface over it. Weekly reporting went from six hours to thirty minutes.

Jun 2024 — Sep 2024

Data Analyst Intern

Acer · San Jose, CA

• 30+ years of records were scattered across siloed systems, so I created a single data model in PostgreSQL that helped increase data accessibility by 125% for Acer's legal team.
• I helped cut report preparation time for General Counsel by creating automated Tableau dashboards, and I elevated licensing discrepencies through analyzing trends from merger and financial docs.