Content-Aware Sparse Attention for BART

Standard attention scales quadratically with sequence length — existing fixes like Big Bird use hardcoded patterns that don't adapt to input. We're building a mechanism that reads the content and decides which tokens to attend to dynamically.

PyTorch CUDA Transformers DeepSpeed BART

Unified Action Adapter Layer

Building the execution backbone of an AI platform — routing NLP-parsed commands across 4 platform APIs with field-level validation, Redis-backed OAuth, and automated credential management.

Node.js Redis PostgreSQL FastAPI OAuth 2.0

Fine-tuned LLaMA 3.1 8B with LoRA + DeepSpeed on 5K debate transcripts, achieving 87% agreement with GPT-4 judgements. Deployed a self-hosted inference pipeline with 70% latency reduction over baseline.

Python PyTorch LoRA DeepSpeed LLaMA

Multi-agent MCP pipeline orchestrating 5 forensic agents. Maps 20 LLM signals to a 0–100 fraud score with async FastAPI + SendGrid webhook backend dispatching replies in under 30s.

Python FastAPI MCP SendGrid SQLAlchemy

100× speedup over CPU across 5M simulations using CUDA on RTX 4060. Engineered 16-byte memory layouts for coalesced memory access with bit-for-bit CPU/GPU parity validation.

C++ CUDA CMake RTX 4060

Trained a Binary Neural Network in PyTorch to classify "open sesame" from raw audio. Deployed fully on-chip to a DE1-SoC FPGA in C, interfaced with a servo motor for a physical gate.

C Python PyTorch FPGA DE1-SoC

C++ city-scale mapping engine with an adjacency list loading datasets in under 2 seconds. Implemented A* pathfinding with cache-friendly data structures on GTK/EZGL.

C++ CMake GTK/EZGL A*

No-code workflow automation platform on Cloudflare Workers. Managed state for 100+ distributed cron jobs via Durable Objects, with LLaMA-powered auto-generation of workflow steps.

Cloudflare Workers Durable Objects LLaMA