Portfolio / 2026

Software Engineer

Charlottesville, Virginia

A geometric Aegean city aligned with a rising sun

Plate 01 / Ambition needs systems

Engineer · Researcher · Builder

Brett Haas

I build intelligent systems that hold up in the real world.

Full-stack software engineer building reliable, intelligent products across web, mobile, and AI systems.

Explore selected work

01 / Flight log

Trajectory

From research and evaluation to production mobile systems: a path shaped by rigor, iteration, and useful outcomes.

A rising coastal city and lighthouse at dusk

  1. 01

    University of Virginia

    ML Research Assistant

    Inference-time activation-steering research focused on language-model visual bias.

    Nov 2025 – May 2026Charlottesville, VA

    • Built an inference-time activation steering pipeline for LLMs in PyTorch and NNSight, intervening on residual-stream activations across Qwen models with frozen weights.
    • Engineered a filtering pipeline over 616K+ COCO captions using lexical scoring, next-token probability disparities, and RMSE-based layer selection.
    • Reduced RMS next-token bias 28.9% while preserving output quality across 1,000 samples.
    • PyTorch
    • NNSight
    • Activation Steering
    • Interpretability

  2. 02

    Scale AI

    GenAI Technical Advisor Intern · SEAL

    Frontier-model safety, evaluation, and agent reliability work on the Scale AI SEAL team.

    Jun 2025 – Dec 2025Remote, USA

    • Red-teamed frontier LLMs in Scale AI SEAL, identifying jailbreaks, unsafe behaviors, prompt-injection risks, and agent security failure modes.
    • Produced RLHF-style code evaluation data for complex software-engineering and competitive-programming tasks.
    • Improved system prompts and tool use in agentic workflows to increase model safety, consistency, and multi-step reliability.
    • LLMs
    • RLHF
    • Model Safety
    • Evaluation

  3. 03

    Refraction Innovation Hub

    Software Engineer Intern

    Cross-platform nutrition and food-recognition product development across mobile, AI, and cloud systems.

    Jun 2025 – Aug 2025McLean, VA

    • Architected a cross-platform food recognition app in TypeScript and React Native with OpenAI multimodal APIs.
    • Deployed authentication through AWS Cognito with storage on AWS RDS and Azure SQL across iOS and Android.
    • Improved load time 55% and reached 99.5% crash-free sessions through memory optimizations and lazy-loading monitored with AWS CloudWatch.
    • TypeScript
    • React Native
    • OpenAI
    • AWS

  4. 04

    University of Virginia

    B.S. in Computer Science

    Systems, cybersecurity, software engineering, and machine-learning research.

    May 2026GPA 3.7

02 / Field studies

Selected work

Three builds across computer vision, real-time collaboration, and language-model research. Choose a panel to unfold the complete case study and verified outcomes.

Case 01

React Native · TensorFlow Lite · Flask · WebSockets · AWS EC2

Court Vision

Computer vision basketball analytics platform for shot tracking and player movement insights.

A geometric arena with analytical trajectory arcs

Brief

Turn basketball footage into practical, coach-friendly signals for reviewing shots, movement, and possession.

Approach

Extract event-level patterns from video, then organize the results in a focused visualization layer for fast game review.

Technical focus

Computer-vision analysis, event pipelines, movement trends, and readable sports analytics.

View project results
89%
shot-detection accuracy
<200ms
on-device inference latency
Repository

Case 02

TypeScript · React · Next.js · Node.js · PostgreSQL · WebSockets

Beat Stream

A real-time collaborative music production platform built for low-latency multi-user sessions.

Coastal architecture crossed by rhythmic signal ribbons

Brief

Let musicians produce together in the same browser-based session without collaboration getting in the way.

Approach

Coordinate multi-user changes through WebSockets, operational transformation, and CRDT-based state management.

Technical focus

Low-latency synchronization, complex PostgreSQL session models, and a component-driven Next.js interface.

View project results
20+
concurrent collaborators
<100ms
real-time sync latency
Repository

Case 03

Python · PyTorch · NNSight · LLM Evaluation

Vision Bias Steering

LLM steering experiments for shifting model outputs between spatial and descriptive language.

A labyrinth observatory with two controlled light paths

Brief

Study whether inference-time steering can shift captions between spatial and descriptive language without retraining model weights.

Approach

Train and validate steering vectors, run local and multi-model sweeps, then compare language behavior through evaluation and plotting utilities.

Technical focus

PyTorch experimentation, NNSight interventions, reproducible sweeps, and language-model evaluation.

View project results
616K+
COCO captions filtered
28.9%
RMS next-token bias reduction
<0.1%
output degeneration
Repository

03 / Craft

The craft behind the flight.

My work sits where software engineering, machine learning, and product judgment meet.

I like difficult systems with visible stakes: steering model behavior, testing frontier agents, and shipping cross-platform products whose performance can be measured—not merely described.

A cliffside workshop with sculptural wings

Research systems

Production engineering

Human-centered interfaces

Research systems Production engineering Human-centered interfaces
A calm sunrise horizon between distant mountain ridges

04 / Next horizon

Keep building.

I’m interested in ambitious engineering teams working across intelligent systems, reliable products, and applied research.

Email Brett