Selene Gisela
portfolio · est. 2026

Hi, I'm Selene Gisela, I turn curiosity into research and research into things that work.

Engineer turned AI builder. I spent years designing physical car parts before falling for language models, and now I build agentic systems that pull structure out of noisy real world text. Currently fascinated by how far you can push an LLM before it needs a human.

about

I started out in mechatronics, designing and releasing physical car parts at Stellantis. Somewhere in there I got stuck on a question I couldn't let go of: is a control model with feedback really any different from a machine learning model? Both take in signals, correct themselves, and get better at hitting a target. Chasing that question is how I ended up in AI.

It led me to a master's in Data Science and Computational Linguistics at UBC, where the thing that hooked me turned out to be language: the messiest, most ambiguous signal of all. These days I am exploring and learning about LLM powered systems and agentic pipelines: the kind of work where you take noisy, real world text and coax something structured and reliable out of it.

I love problems that sit between disciplines, where engineering discipline meets human intent, and where the hard part isn't the model but figuring out what "correct" even means. I like making systems that are honest about what they know, and useful because of it.

Based in
Vancouver, BC
Focus
NLP, Supervised Learning, Unsupervised Learning, Data Engineering
Portrait of Selene Gisela
experience

Where I've worked

Apr 2026 — Jul 2026

Jeppesen ForeFlight

NLP Engineer — Capstone Project

  • Built and shipped an end to end LLM extraction pipeline that turns noisy transcripts into 11 structured fields, replacing a manual reporting process, covering data preparation, prompt engineering, evaluation, and productionization.
  • The extraction engine is a three layer extraction cascade regex → dictionary lookup → LLM fallback that parses structured emergency event fields from noisy ASR/OCR/subtitle ATC transcripts in a Databricks pipeline
  • Designed a LLM agent with three stages(Spotter, Classifier, Linker) with prompt chaining for entity recognition, role assignment, and relationship extraction.
  • Deployed it as a FastAPI REST service on Databricks Apps against a live LLM serving endpoint, owning architecture through production.
  • Added RAG grounding plus validation against a 5,000+ record reference dataset to cut hallucinations and catch bad output before release.
May 2022 — Aug 2025

Stellantis

Design Release Engineer - Product Engineer

  • Collaborated on 5 cross-functional teams and 10 suppliers through a new vehicle program, translating rough requirements into technical specs.
  • Analyzed test and warranty data for root causes and drove fixes through formal change control, closing change notices 20% ahead of schedule.
  • Wrote formal engineering documentation and validation plans where precise, unambiguous wording carried contractual and safety weight.
May 2019

Beautiful Patterns (MIT & Tec de Monterrey)

Teaching Assistant — Volunteer

  • Supported computational thinking exercises (HTML, C++, Python) in a program for women in STEM.
selected projects
data engineeringNLP

Capstone Project

End to end LLM extraction pipeline for automated reporting.

View project →
researchpytorch

Membership Inference Attack

Privacy attacks that detect whether specific samples were in a model's training data, including a low-confidence-token method that was the strongest single attack on a fine-tuned LLM plus contrastive-margin attacks for the multimodal setting.

View project →
RAGdata engineeringnlp

RAG Cookbook

End-to-end retrieval-augmented generation system that parses a cookbook into structured data, embeds it for semantic search, and answers grounded questions with LLM routing and a recipe scaling action.

View project →
nlplow resource

AI Foreshadowing

1M word annotated corpus with BiLSTM and LR models via transfer, multitask, and few-shot methods for detecting narrative foreshadowing, deployed as a Dockerized FastAPI + Streamlit app.

View project →
classificationpytorch

Rating Predictions

5 class sentiment classifier for 35K Yelp reviews that benchmarks a TF-IDF + Logistic Regression baseline against a tuned PyTorch neural model, packaged as a tested, modular pipeline with a CLI.

View project →
classification

Default Prediction

Gradient-boosted classifier predicting credit card payment default from account and repayment history features.

View project →
classificationnlp

NLI Classifier

NLP pipeline that classifies English learners' native-language family from their writing, reaching a 9-point lift over baseline, using three interpretable feature families and a grid-searched decision tree.

View project →
toolbox
NLPSupervised LearningChange ManagementDeep LearningUnsupervised LearningContext EngineeringA&DSResearchData VisualizationData EngineeringDashboardsCorpus LinguisticsMachine TranslationComputational SemanticsComputational Morphology
paper notes

Notes, highlighted

Notes on papers I'm reading or topics of my interest; what they claim, what convinced me, what didn't. Written for future me, shared in case it's useful to you.

№ 001 2026-04-22 topic · Low resource languages

Zapotec as a low resource language in Machine Learning

On: "Flash Talk", Selene Morales, 2026

What is a low resource language, and what does Zapotec — one of Mesoamerica's oldest writing traditions — teach us about the limits of NLP?

Read note →
№ 002 2026-04-12 topic · Transactions on Affecting Computing

Empathy in technology

On: "Empathy by Design: The Influence of Trembling AI Voices on Prosocial Behavior", Fotis Efthymiou and Christian Hildebrand, 2024

Notes on how trembling AI voices shape empathy, perception, and prosocial behavior in conversational AI.

Read note →