Hey, I'm Alex

In December 2024, I built Antidote Intelligence to help find suspicious examples in training datasets before they affect a model.

Anthropic, the UK AI Security Institute, and The Alan Turing Institute later published a large scale poisoning study that makes that concern feel a lot less speculative. Their result: in their experimental setup, as few as 250 malicious documents were enough to introduce a backdoor across models from 600M to 13B parameters.

What caught my attention

The number that caught my attention was 250 documents. Their finding suggests the absolute number of poisoned documents can matter more than the percentage of the corpus, at least for the narrow backdoor they tested.

What Antidote was trying to do

Antidote was not trying to solve all model security. It was a dataset inspection tool: look at examples before they become model behavior, generate hypotheses about suspicious content, and make data quality visible enough for a human to investigate.

Read Anthropic's research post.

AeVAA is a research project where I worked on tracing how a system assigns sentiment to different people or things in a sentence.

Aspect based sentiment analysis usually tries to predict whether a sentence is positive, negative, or neutral toward a target. That is useful, but it often hides the path from text to judgment. AeVAA takes a different route: split the problem into modules, keep intermediate artifacts, and use survey derived formulas to explain how sentiment moves between entities.

The core idea

Instead of asking one model for one answer, AeVAA builds a trace. It extracts clauses, resolves entities, identifies relationships, constructs a graph, and then calculates valence aware sentiment over that graph. The model can still use black box components, but the system around them exposes what each component contributed.

What I wanted to understand

I wanted to understand who a sentence is actually positive or negative toward, and be able to follow how the system reached its answer. In a sentence like "the person was bad, but the child was good," one overall sentiment score does not tell me much. I care about who the sentiment is aimed at and why it changed. That is what drew me to this problem, especially for media bias and longer stories where framing changes how we read someone.

What we built

• A modular pipeline for constituency clause extraction, entity/coreference resolution, relation and modifier extraction, graph construction, and sentiment aggregation.
• A human annotation study with 36 participants and 3,900+ sentiment judgments across action, association, ownership, and temporal aggregation cases.
• Survey fitted formulas for action, target, association, ownership, and aggregate sentiment dynamics.
• Explanatory traces that show where errors came from instead of only reporting a final label.

Results

The fitted formulas explained roughly half of the variance in pilot sentiment judgments. On SemEval 2014, AeVAA reached 78.58% restaurant accuracy and 68.52% laptop accuracy. Its accuracy was below state of the art DeBERTa systems. The useful part for me was being able to trace a wrong answer back through the pipeline and see which step contributed to it.

Maldemic is a stochastic disease spread simulator that turns SIR equations and city mobility into a live 3D globe. I like projects where the math becomes something you can inspect with your eyes.

Data flow

Python computes population movement with a Markov matrix, updates local susceptible/infected/recovered states, then passes the evolving state into a Godot visualization.

What went into it

• Markov chain mobility between cities
• SIR disease dynamics for local spread
• Population cleanup to keep totals consistent
• 3D globe rendering for real time visual feedback

The project won 2nd place at BLOOM Hackathon and received grant support for neural network forecasting work.

I demoed Antidote Intelligence at Continue DX, showing a content aware data poisoning detection system for ML training datasets. The demo focused on reviewing the contents of a dataset before training.

Why I started with the data

If bad examples or poisoned content can shape its behavior, I wanted a way to inspect those examples before training. That was the idea behind Antidote: help me see what looks suspicious in a dataset and understand why it deserves a closer look.

What I built

• Multi agent review pipeline for suspicious dataset content
• Hypothesis generation and validation around poisoned examples
• Content aware checks instead of only metadata based filtering
• Reports on suspicious examples for a person to review

This one is more of a journal entry than a project breakdown.

I transferred to USF in 2024 after UC Merced because I wanted to be in San Francisco. Merced felt too far away from the people and companies I wanted to learn from. I wanted to make a name for myself, be around real builders, and learn from people actually working in tech.

Before I even got to USF, I applied to more than 100 jobs. That search eventually turned into a Data Analyst / Web Intern role, where I worked on enrollment analytics, SLATE data, admissions event cleanup, and prospective student web updates. A lot of the work involved cleaning up records and updating pages for prospective students.

Where sentiment came in

Later, at the MAGIC Lab, I worked on AeVAA, an explainable sentiment project. I wanted to understand how the system reached a sentiment score, so I worked with graphs, relation extraction, and formulas fitted to survey responses.

In 2023, I was the secretary of the UC Merced Game Dev Club. I helped plan events, message people, book rooms, and welcome students who had never made a game before.

We hosted a showcase where people brought in games they had been building, talked through what worked, and compared notes on controls, art, music, level design, and keeping their games small enough to finish.

Game jams and mixers

We also hosted a game jam that produced several student projects. I liked seeing people meet, form teams, and finish a game together. I also helped host a mixer for students who wanted to get started with game development but did not know who to work with yet.

That year also overlapped with my first day as a Data Analyst Intern at SATAL UC Merced. I was using survey data and student feedback to improve enrollment support, course experiences, and faculty decision making.

Presenting at Fresno State was a chance to share the survey work beyond UC Merced and show what students were saying about their courses.

In 2022, I had just started college and I could not get an internship yet. I was looking for work I could do over the summer.

So I worked with ACME Builders. Some of it was construction work. Some of it was office work: scanning accounting documents, payroll records, and old archives so they were easier to store and reference. I got to see both the construction work and the paperwork that went with it.

At the time, it was mostly a way to spend the summer working while I figured out what I wanted to do next.