Brandon
Collins
CS & Physics — Auburn University
Undergraduate researcher focused on Computer Vision and Human-AI Interaction. First author on SketchVLM (EMNLP 2026 Main) and first author on the PSR dataset (CVPR 2025 Workshop & WACV 2026). Vice President of Auburn's ACM chapter. Former Quantitative Strategist intern at Goldman Sachs.
Understanding Generative AI Capabilities
in Everyday Image Editing Tasks
We propose the PSR dataset — 305,000 real image-edit requests web-scraped from Reddit — and evaluate state-of-the-art multimodal models including GPT-4o and Gemini, showing AI can complete 33% of real-world editing requests.
- Designed and backtested adverse-selection signals, reducing negative markouts for electronic market-making in interest-rate products.
- Analyzed dependencies across client analytics jobs to identify critical-path bottlenecks causing processing delays and SLA breaches.
- Presented markout analysis and bottleneck dashboard to traders, desk strats, and engineers.
- First author of SketchVLM (accepted at EMNLP 2026 Main), a training-free framework for generating editable visual explanations with VLMs, improving accuracy by up to +28.5 points and annotation quality by up to +48.3% over image-editing and fine-tuned models.
- Led manuscript writing, implemented Python inference and evaluation (100+ GitHub stars), created benchmarks, and built the user demo.
- First author of Understanding Generative AI Capabilities in Everyday Image Editing Tasks (accepted to CVPR 2025 Workshops and WACV 2026); introduced PSR, a dataset of 305K edits derived from real user image-edit requests, finding SOTA multimodal models complete 33% of requests.
- Built an automated Python and Playwright pipeline to submit 1,000+ image-editing requests through ChatGPT and capture generated outputs before image-editing API access was available.
- Computing Research Association Undergraduate Research Award — Honorable Mention
- Coached 40+ students across 8 STEM courses (Calc III, Linear Algebra, Physics, etc.), improving understanding and exam performance.
- Streamlined JIRA status updates for 100+ technology findings and created SOPs and remediation plans.
- Raised executive awareness of 400+ application-security vulnerabilities to protect 20K+ financial advisors.
- Built a C++20 engine reconstructing Nasdaq TotalView-ITCH order books for all 8,700+ securities (368M msgs/day) at 13.9M msg/s (p50 62 ns), 4.3× faster than an STL baseline, verified by checksum against an independent Python implementation.
- Built a queue-aware market-making backtester and found that thin-side fills are 1.1 bps more adverse only in large-tick stocks.
- Coach 30 students in advanced algorithms, data structures, and contest strategy for ICPC.
- 1st place — ICPC 2025 University of West Florida Regional D1
- 1st place — ICPC 2023 University of West Florida Regional D2
- Manage funds, social events, and company tech talks. Oversee four CSSE clubs serving 300+ students: Competitive Programming, Technical Interview, Web Development, and Ethical Hacking.
- Mentored 50+ members on algorithms, technical interviewing, and the recruiting process.
Open to research collaborations, opportunities, and conversations.
blc0063@auburn.edu