Hey, I'm Alex

In December 2024, I built Antidote Intelligence around a simple belief: training data is infrastructure, and poisoned examples can become model behavior if nobody inspects the dataset early enough.

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.

Why this reinforced the project

A lot of people think poisoning only matters if an attacker controls a meaningful percentage of the training set. The Anthropic result challenges that. 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.

The larger lesson

This is the same theme as my recommender and ABSA work: AI systems can be powerful without becoming completely opaque. If model behavior depends on messy upstream data, then inspection, provenance, and explainable intermediate artifacts are not extras. They are part of the system.

Read Anthropic's research post.

AeVAA is a research project about a question I keep coming back to: machine learning and AI are powerful, but can we build systems where the important reasoning stays inspectable?

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.

Why this matters

Document-level sentiment can miss the point. In a sentence like "the person was bad, but the child was good," the total sentiment is not enough. The meaningful question is who the sentiment is aimed at and why it changed. That becomes even more important for media bias, long-form narrative, and texts where framing matters.

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. It did not beat state-of-the-art DeBERTa systems, but that was not the point of the prototype. The point was to show that a modular, inspectable ABSA system can produce plausible results and make debugging easier.

The bigger theme

I like projects that score well without becoming total black boxes. The goal is not to reject ML; it is to use ML where it helps, then design the surrounding system so people can inspect the evidence, the intermediate state, and the reason a result appeared.

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.

Technical shape

• 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 basic pitch: before we argue about the model, let's look harder at the data we fed it.

Why this matters

Training data quality is one of those problems that hides until it becomes expensive. Bad examples, poisoned content, or subtle distribution weirdness can leak into model behavior long before anyone notices.

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 aimed at making data issues inspectable, not magical

I am interested in this space because it treats data quality as infrastructure. The model gets the attention, but the dataset is where a lot of the story starts.

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. It was not glamorous, but it taught me something important: useful software usually starts as messy data, weird processes, and people who need better tools.

Where sentiment came in

Later, at the MAGIC Lab, I worked on AeVAA, an explainable sentiment project. The technical version is about aspect-based sentiment analysis, graphs, coreference, relation extraction, and survey-derived formulas. The personal version is simpler: I was trying to understand how a system could make a judgment without hiding the reasons.

That thread shows up in a lot of my projects. Recommenders should explain why a game matches. Sentiment systems should explain who a sentence is about and why the score moved. Data poisoning tools should show what suspicious examples look like before they become model behavior.

The pattern

I like AI systems, but I do not like when the answer is the only artifact. The projects I keep coming back to are the ones where the intermediate state matters: vectors, tags, traces, formulas, records, examples, and the evidence behind a prediction.

USF was where that started to become a theme instead of a coincidence.

In 2023, I was the secretary of the UC Merced Game Dev Club. A lot of the work was not glamorous: planning, messaging people, getting rooms, keeping events moving, and making sure students felt like they could show up even if they had never shipped anything before.

We hosted a successful showcase where people brought in games they had been building, talked through what worked, and got to see other students care about the same weird problems: controls, art, music, level design, scope, and how to make a tiny idea feel playable.

Game jams and mixers

We also hosted a game jam that produced some genuinely cool student projects. The best part was watching people form teams quickly and make something real under a deadline. 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. Looking back, both roles were about the same thing: turning scattered student energy into something organized enough that people could act on it.

That presentation mattered to me because it made the internship feel real. We were not just cleaning survey data for a class assignment; we were turning student perspectives into something instructors could discuss, revise around, and bring back into their courses.