Experience
-
Jun 2026
– Sep 2026Mibanco — Credicorp
Artificial Intelligence Intern · Lima, Peru 12°03′S
- Problem
- Recorded customer audio had to become structured, usable data for production workflows.
- Built
- A parallelized audio-ingestion and speech-to-text pipeline on Databricks with Azure AI Speech, plus a prototype LLM interpretation layer over the transcripts.
- Result
- Cut end-to-end runtime from 7 min per request (20 min under load) to a stable 5 min. Transcription was partially adopted on real customer calls at Peru’s largest microfinance bank; rollout paused for other priorities.
- Stack
- Databricks · Python · SQL · Azure AI Speech
the four-stage Databricks workflow, parallelized
- audio
- transcription
- encryption
- LLM feedback (prototype)
- email delivery
- Built a parallelized audio-ingestion and speech-to-text pipeline on Databricks (Python, SQL) using Azure AI Speech, at Peru’s largest microfinance bank; partially adopted on real customer calls, with the rollout paused for other priorities
- Cut end-to-end pipeline runtime from 7 minutes per request, degrading to 20 minutes under concurrent load, to a stable 5 minutes independent of load — by parallelizing the concurrent work and restructuring the code rather than adding compute
- Prototyped LLM-based interpretation of transcribed speech, evaluating Azure AI models on output quality, cost, and workflow fit
- Closed the internship by validating the team’s proposed single sign-on architecture — one Microsoft Entra ID login shared by two web apps, one embedded in the other — with a runnable proof of concept (Next.js front ends, a mock identity provider, two APIs validating the same JWT) and attack and isolation tests; surfaced missing tenant, scope and framing checks and presented it as the design to carry toward production
- Coordinated the AI team’s project planning while automating routine engineering tasks with LLM-assisted tooling
-
Nov 2025
– presentTRUE (Trustworthy and Responsible) AI Lab — Oregon State University
Undergraduate Researcher — LLM Reliability & Safety · Corvallis, OR 44°34′N
- Problem
- Hardware bit-flips can silently corrupt LLM inference — how robust are models, and what defends them?
- Built
- Robustness evaluations on the lab’s fault-injection harness, run at scale on an A40 Slurm cluster.
- Result
- Co-authored a cross-scale resilience manuscript, now under review. Separately, ran a 42-cell sweep: 7 defense configurations × 2 attack classes × GSM8K, DROP and TriviaQA.
- Stack
- PyTorch · Hugging Face · Slurm
- Evaluate robustness of Llama-3-8B to hardware bit-flip faults using the lab’s fault-injection harness (PyTorch, Hugging Face), run at scale on an A40 Slurm HPC cluster
- Co-authored manuscript on cross-scale model resilience across 8B–32B target models, now under review; contributed experiments, benchmarks and the accuracy-versus-fault-tolerance trade-off analysis
- Ran a separate 42-cell study benchmarking 7 fault-tolerance defense configurations against 2 classes of adversarial bit-flip attacks (stealth and damage) across GSM8K, DROP and TriviaQA
-
Jan 2026
– presentOregon State University Foundation
Digital Technology Student Assistant · Corvallis, OR 44°34′N
- Problem
- Operational datasets needed to become institutional reports, and production web content needed upkeep.
- Built
- Data visualizations and summary reports from operational data; production web pages maintained in Sitefinity CMS.
- Result
- Reporting and web content delivered for institutional use.
- Stack
- Data visualization · Sitefinity CMS
- Build data visualizations and summary reports from operational datasets for institutional reporting, and analyze engagement metrics with Matomo and Fundmetric for the Foundation’s digital properties; maintain production web pages in Sitefinity CMS
-
Jul 2024
– Aug 2024Universidad de San Martín de Porres
Python Course Assistant (volunteer) · Lima, Peru 12°03′S
- Problem
- Students needed hands-on help with Python assignments and debugging.
- Built
- One-on-one mentoring, rubric-based grading, and recurring study sessions.
- Result
- Led 5+ study sessions and graded coursework against detailed rubrics.
- Stack
- Python
- Mentored students on Python assignments and debugging, graded with detailed rubrics, and led 5+ study sessions