Built four Claude-based research tools for equity and credit analysts, including a batch Porter’s Five Forces pipeline that automated manual reviews at 99% concordance with senior analyst baselines, cutting turnaround by weeks.
Automated quarterly research updates with a pipeline that diffs new SEC filings and earnings materials against prior snapshots and records what changed for public-equity and credit review.
Shipped an AlphaSense-fed updater that refreshes existing analyst workbooks with source-attribution links and validation checks, and a screener that narrows the full ticker universe to a short list for deeper research.
Built a New York State online pre-licensing course to DMV DTP-402 rules (Next.js, TypeScript, Firebase): server-enforced 270-minute seat time, identity checks down to keystroke dynamics, a three-strike lockout, and electronic completion reporting to the DMV. Awaiting DMV approval.
Deploys gated on 70+ compliance, end-to-end, and security test suites (Docker, GitHub Actions, AWS App Runner); idempotent Stripe payments; S3 Object Lock retention; daily Firestore backups to S3.
October 2022–August 2023, May–August 2024New York, NY
Co-authored a paper on smartphone heart-rate-variability baselines for concussion in 752 adolescent athletes (Med. Sci. Sports Exerc.; ACSM 2025); built ML pose tracking from video and quaternion-based iPhone motion tracking.
Built MOSS-Image Turbo: 6× faster than the official 50 steps (12.4 min to 119 s, M4 Pro) without the stamped textures of community few-step (DMD) adapters. Traced them to repeated distilled steps mid-trajectory; fixed it with trajectory-regression training on teacher-relabelled targets; passed a pre-registered test on 32 held-out prompts.
Ported OpenMOSS’s 5B bilingual text-rendering DiT, MOSS-Image, to pure MLX, exact against official PyTorch (69.7 dB end to end in fp32; official seeds reproduce official images). The 8-bit tier peaks at 9.3 GB vs. 38.7 GB for mlx-vlm’s port, with OCR recall identical to bf16. Also ported FLUX 3 Image Klein (cosine 0.9999 vs. BFL).
Shipped both ports as pip-installable packages with command-line tools (moss-image-mlx, flux3-mlx).
Extracted Qwen3.8-27B’s multi-token-prediction head as a speculative-decoding drafter (up to 1.89× faster on coding prompts) and released 6-bit MLX weights with a full-precision vision tower; 7,700+ downloads across both.
Distilled Ideogram 4 for MLX by folding classifier-free guidance into one branch and cutting steps: 2.9× faster at 800×992 on an M4 Pro (250 s to 86 s), quantized to 18.6 GB to run on a 48 GB Mac.
Designed a full-factorial eval of how image models respond to filmstock, lighting, and scan language (288 images, 195 features; PCA, MANOVA, PERMANOVA). A September replication found the newer model 24% less responsive to filmstock, 14% to lighting, and 37% to scan language (Holm-adjusted p < .001). 31-page write-up.
Curate an AI-image archive that keeps each image’s model, prompt, and workflow; built its Builder, which compiles camera, lens, film, and lighting choices into prompts for five model formats.
