SOFTWARE ENGINEER · AI BEHAVIOR RESEARCH

I engineer with AI the way behavior is studied: operationally defined, measured, and changed one variable at a time.

ABSTRACTSoftware Development Engineer working across the full stack in Next.js, TypeScript and Azure. My path into engineering ran through seven years as a Registered Behavior Technician, implementing ABA programs and AAC systems for nonverbal children, and that training in operational definitions, measurement and single-subject design still shapes how I build with, and evaluate, AI systems. Independently, I run Aurora, a longitudinal study of what language and vision models produce when given a canvas and no instructions.

Degree: B.A. Computer Science, CU Denver, 2026
Work: Remote
Code: github.com/elijahsylar
Research: Aurora ↗

1Method

Agents generate a substantial share of my code. Reliability comes from the system that governs them: a fixed sequence of specification, review and verification that every change passes through, regardless of its author.

1.1Specification. Each change begins as a written statement of scope, constraints and invariants: what must change, and what must not.
1.2Reconnaissance. An agent surveys the existing codebase without write access, so planning is grounded in what the system actually does rather than what it is assumed to do.
1.3Decision record. Architecture, alternatives considered and known risks are documented and approved before implementation begins.
1.4Gated implementation. Work proceeds in discrete checkpoints, each requiring verifiable evidence before the next is authorized.
1.5Independent review. A separate agent with no shared context audits the complete session for assumption drift, unhandled edge cases and security defects.
1.6Accountable sign-off. I review every diff and finding personally, and the original plan and final decisions are preserved in the pull request as a durable record.

2Work

2.1 · INDEPENDENT RESEARCH · SINCE 2025

Aurora

An experimental platform that presents local and API-hosted models with a canvas, a constrained command vocabulary and no task instructions, then logs each action alongside the model's own stated reasoning. The research question is what a model produces when given the opportunity, but not the directive, to create.

3,152 sessions · 100,562 logged thoughts · dataset from March 2026

Research page ↗Live dashboard ↗

2.2 · SOLE ENGINEER · PAUSED

LumpyLP

A community-ranked music discovery platform, designed as an alternative to algorithmic curation. Its ingestion pipeline normalized approximately 448,000 Creative Commons and public-domain tracks from more than eleven heterogeneous sources through source-specific resolvers, a centralized license-compliance gate and a unified catalog.

Case study →

2.3 · CURRENT ROLE

Software Development Engineer

Full-stack ownership of a production application, spanning its data model, server-rendered React interface, infrastructure as code, and the language-model integrations between them. I joined as an intern and have served as the application's primary developer since; the method described above is how that work is delivered.

Next.js · React · TypeScript · Node · Azure (Cosmos DB, Entra ID, Container Apps, OpenAI) · Terraform · GitHub Actions · Vitest

2.4 · OPERATIONS

Library system

Sole database administrator for a production Koha integrated library system serving more than 200 users, sustaining 99.9% uptime across 18 months.