Summary
AI/ML engineer specializing in applied GenAI systems, multi-agent orchestration, and GPU-level performance work. Takes ambiguous problems to shipped products as sole owner, with depth across computer vision, NLP, real-time systems, and model interpretability.
Experience
Autolog – AI-Powered Knowledge Acquisition Platform
- Architected and shipped a 130K-line production AI platform as sole engineer with full ownership — designed to bridge the gap between low-response-rate structured surveys and costly in-person interviews. Built on multi-agent orchestration with a dual-process architecture (Analyst reasoning + Conversationalist generation), DAG-based real-time processing, streaming speech-to-text (Deepgram Flux), a WebRTC relay, session persistence, and artifact storage on Cloudflare R2
- Coding Harness: accelerated AI-assisted development with a worktree-scoped, ephemeral observability stack (Loki/Promtail/Prometheus via Docker Compose), Chrome DevTools Protocol automation (screencast recording, console capture, audio injection), structured Pino logging, and a custom skill library for navigation, audit, and implementation-review
Client Engagements — Pre-Sales · Technical Advisory and Delivery
- Bookmarked (Content Moderation): Fine-tuned LlamaGuard 3 8B with LoRA adapters for a client moderation product; evaluated 5 candidate datasets by annotation methodology and coverage, then selected a training mix targeting implicit-pattern detection rather than explicit violations alone. Built an nltk text-extraction and sentence-tokenization pipeline to classify the full client dataset, and delivered a technical recommendations report with model comparisons and a deployment path
- Fit a ~50GB model onto a single 24GB L4 GPU — removing the need for multi-GPU infrastructure — by building CUDA memory-profiling utilities (
torch.cuda.memory snapshots) and tuning DeepSpeed ZeRO Stage 3 (persistence thresholds, prefetch bucket sizes, gradient checkpointing, pinned-memory offloading); instrumented training with NVTX markers for GPU-timeline analysis
- Innova (LLM Agent): Built an auto-repair instruction agent that aggregates YouTube transcripts and other automotive sources into step-by-step repair guidance, with a Python web app for internal demos and LangChain structured-output parsing with retry logic. Deployed on AWS with nginx/gunicorn; earned Google Cloud Professional certification during the engagement
- Hannagriff (Computer Vision): Built an end-to-end scantron-grading pipeline — MSER extraction, DBSCAN clustering, vector-field-guided grid drawing, and CNN classification trained on 20K hand-labeled samples (evaluated by macro-F1, with oversampling to address class imbalance) — converging after iterating through 5+ algorithmic approaches. Developed an agentic auto-research loop to refine data-labeling heuristics; built with Python, OpenCV, scikit-learn, and Shapely; deployed on AWS EC2 with S3 integration and handed off to the client
- Built an interpretability pipeline (LIME, CAM, HiResCAM) for a ResNet50+Attention gestational-age model on prenatal ultrasound video, enabling clinicians to understand the model's reasoning behind prenatal-care decisions; generated heatmaps across 300+ clinical exams for sonographer review
- Forked scikit-image to modify quickshift++ segmentation in Cython for domain-specific superpixel constraints, then trained ridge-regression interpretability models (NumPy) on the resulting superpixel features
- Prototyped 3 architectures for multiple-gestation classification (attention retrained on the GA ResNet, a GA+YOLO feature ensemble, and MobileNet with custom attention); diagnosed overfitting and pivoted to feature-engineered classical ML backed by a SQL feature store with automated extraction, with full experiment tracking (metrics, hyperparameters, git provenance), delivered as a formal scientific presentation
- Designed (but did not build) a Stable Diffusion fine-tuning pipeline for synthetic ultrasound-video generation across two clinical study datasets
- Built a pytest test suite for a data ETL pipeline processing smartphone data
- Reverse-engineered a legacy, linear-algebra-based statistical model for military airdrop systems — rebuilding the understanding of an undocumented algorithm from probability and linear-algebra fundamentals, and earning the trust of the team
Projects
Abstruct: An AI-enabled UX for memorization and concept engagement that turns ideas into content — GenAI-powered Socratic recall using FSRS scheduling, real-time voice input (Deepgram Nova-2), continuous LLM evaluation, and tree-structured conversation branching; currently designing a handheld companion device (Raspberry Pi Zero 2W, e-ink display, camera OCR, MEMS microphone) for screen-free capture of primary-source material with user commentary
Fuel-Code: Developer-productivity platform built on a deep understanding of Claude Code internals — models the full agent lifecycle (sub-agents, teams, skills, worktrees, permission modes) to preserve session transcripts and let developers reuse and iterate on productive sessions
Arxiver: RAG-LLM research tool that spins up expert chatbots in user-specified domains
CUDA Deep Learning: From-scratch neural network in CUDA and C++ — hand-written kernels for matrix ops, activations, and losses; forward + backward passes on-device; multi-architecture builds via Makefile fat binaries targeting Volta (sm_70) and Pascal (sm_61)
SimpleProcessor: Designed and simulated a custom CPU in Verilog with a C++ test harness — implemented the full instruction fetch/decode/execute pipeline, register file, and ALU, working hands-on across the hardware-software boundary
MemoryAllocator: Custom memory allocator in C — free-list management, sbrk/mmap-backed regions, and fragmentation analysis, validated by a dedicated test harness; performance-critical low-level memory work
NUFileSystem: FUSE-based filesystem in C — block allocation, inode tables, and directory-tree traversal implementing the POSIX interface, with a full test suite covering the syscall surface
Education
Northeastern University — BS Mathematics, Cum Laude, 3.70 GPA
Eta Kappa Nu — EECE Honors Society
Coursework: Machine Learning, Probability, Linear Algebra, Real Analysis, Group Theory, Algorithms, Computer Systems, Theory of Computation
Certification: Google Cloud Professional (Oct 2025)
Skills
Languages: Python, TypeScript, SQL, C/C++, CUDA, Verilog/HDL, Go
AI/ML: PyTorch, PyTorch Lightning, LangChain, Anthropic SDK, MCP, DeepSpeed ZeRO, LoRA/PEFT, OpenCV, scikit-learn, NumPy, Cython, nltk, Deepgram
Systems & Performance: Custom CUDA kernels, multi-architecture GPU builds (Volta/Pascal), manual memory management, FUSE-style filesystems, processor / HDL design, real-time streaming pipelines, low-latency WebSocket / WebRTC protocols
Infrastructure: Docker, Redis, PostgreSQL, AWS (EC2/S3/EBS), Supabase, Cloudflare R2, nginx, Railway, Vercel
Web & Real-Time: React, Hono, Express, Bun, WebSocket, WebRTC, Chrome DevTools Protocol, Server-Sent Events
Observability & Tooling: Prometheus, Loki/Promtail, Pino structured logging, NVTX, torch.cuda.memory profiling, Turborepo, Drizzle ORM, Git, pytest
Interests: Guitar, staying active, time with loved ones.