I'm a multimodal and forward-deployed AI engineer. I port models to MLX, design image evals, and ship full-stack products.
I'm a multimodal and forward-deployed AI engineer. I port models to MLX, design image evals, and ship full-stack products.
Summer AI AnalystNew YorkMay–August 2026
Built four Claude-based research tools for equity and credit analysts. A batch Porter’s Five Forces pipeline automated manual reviews at 99% concordance with senior analyst baselines, cutting turnaround by weeks.
Co-Founder / DeveloperNew YorkApril 2025–present
Built a New York State online pre-licensing course to DMV DTP-402 rules: server-enforced 270-minute seat time, identity checks down to keystroke dynamics, electronic completion reporting to the DMV. Awaiting DMV approval.
Mount Sinai Hospital, Abilities Research CenterResearch AssistantOctober 2022–August 2023, May–August 2024Co-authored a paper on smartphone heart-rate-variability baselines for concussion in 752 adolescent athletes; built ML pose tracking from video and quaternion-based iPhone motion tracking.
Models on Apple Silicon6.2× faster MOSS-Image, in 4× less memory · 2.9× faster Ideogram 4 · 7.7K downloads of a 27B modelOpen-weight MLX releases of Qwen3.8-27B, Ideogram 4, FLUX 3 Klein and MOSS-Image: ported, quantized, distilled and checked against the official pipelines on one Mac.
Filmstock ResponsivenessChatGPT Images 2.5: 24% less responsive than 2.0 to filmstock, 14% to lighting, 37% to scan languageA full-factorial eval of how image models respond to filmstock, lighting and scan language: 288 images, 195 features, a 31-page write-up.
Gesso ClubUpdated daily · prompts for five model formatsA daily-updated archive of AI images that keeps each image’s model, prompt and workflow, and a Builder that compiles photographic choices into prompts for five model formats.Repeated texture in few-step distilled image models Notes on the MOSS-Image few-step mode; a pre-registered test on SDXL, Qwen‑Image and FLUX.1‑dev2026
Use of Smartphone Application Assessment of Heart Rate Variability in Baseline Concussion Assessments of Adolescent Student AthletesCo-author. ACSM Annual Meeting; Medicine & Science in Sports & Exercise2025
Filmstock responsiveness in image models Sole author. Eval, paper and September 2026 replication2026
Criticality in binary sparse coding Simulation study2026
OpenMOSS’s 5B bilingual text-rendering model on Apple Silicon, exact against the official PyTorch pipeline and 9.3 GB at its peak for a 1024² image. Its --turbo mode makes an image in about 2 minutes instead of 12½.
Open on GitHubBlack Forest Labs’ FLUX 3 Image Klein on Apple Silicon, released the day BFL’s was: text-to-image, editing with up to 10 references and bounding-box layout; it matches BFL’s own render at 35.9 dB.
Open on GitHubAn agent skill that reads a student’s Blackboard courses and keeps a Google Tasks list of assignments, each task opening on the assignment’s own link, refreshed weekly. It shows the proposed updates for approval before the first sync.
Open on GitHubA draft submission to OpenAI’s 16 MB language-model challenge: int6 quantization + zstd, fp16 embeddings, a 3× MLP, 4,096-token sequence length and sliding-window eval.
Open on GitHubTwo ComfyUI nodes (a KJNodes fork) that compile artistic choices into structured Ideogram 4 prompts, with a design note on the controls.
Open on GitHubA local studio for dithering, halftone screens and ASCII, with WebGL2 bloom, halation and chromatic aberration. Exports PNG, WebP, SVG and frame-exact MP4.
Open on GitHubA pure-CSS progressive blur over six layers. A random search over 5,000+ parameter sets cut its banding score from 0.0226 to 0.0086 (2.6× lower).
Open on GitHubCoursework: Advanced Computational Neuroscience, Advanced Statistical Methodology, Microeconomics, Genetics; in progress: Brain-Machine Interfaces, Data Methods for Economists, Developmental Biology.
