I am an Applied AI & High-Performance Systems Engineer specializing in distributed retrieval systems, high-throughput LLM serving pipelines, and low-level algorithmic engineering.
- 🛠️ Applied AI Engineer (Fractional) @ Data Precedes Inc. (Ontario, Canada / Remote) — Architecting enterprise agentic AI workflows, tool-calling ecosystems, and reliable webhook ingestion infrastructure.
- 🚀 High-Efficiency LLM Serving & Agent Protocols — Authoring Anser, the Universal 245K agent microkernel and local serving protocols on consumer hardware (benchmarking 245K-token contexts, ~130 tok/s decode via vLLM + DFlash2 + KVarN integration, powered by syv-ai/qwen38-27b-rtx3090).
- 🌐 Distributed Systems & Hybrid Retrieval — Designing multi-node RAG pipelines combining Tantivy BM25 lexical search, BGE-M3 dense embeddings, and Qdrant gRPC vector storage over encrypted WireGuard meshes.
- ⚙️ Low-Level Systems & Algorithmic Rigor — Engineering zero-dependency assembly game engines, empirical graph benchmark suites, and bounded stream-disk join engines across C++17/20, x86 MASM, and Python.
| System / Repository | Visibility | Architecture & Key Deliverable | Verified Empirical Metric |
|---|---|---|---|
| Anser | 🌐 Public |
The Universal 245K agent microkernel for local LLMs (vLLM, DFlash2, KVarN, MCP) | 245K context in 24 GB VRAM · 78.9% draft acceptance · $0 token cost |
| Geo-Distributed Hybrid RAG | 🌐 Public |
3-node distributed retrieval (FastAPI, Tantivy BM25, BGE-M3, Qdrant gRPC) | Reciprocal Rank Fusion (k=60) · 160ms barrier sync |
| Multi-Task Pathology U-Net | 🌐 Public |
Multi-task pathology U-Net diagnostic benchmark (PyTorch, GradNorm dynamic loss balancing, Macenko normalization) | 26-run paired ablation isolating task interference · Patient-disjoint 5-fold CV |
| Enterprise WhatsApp AI Assistant | 🔒 Proprietary |
Enterprise agentic tool-calling workflows & webhook ingestion @ Data Precedes | Sub-500ms webhook · 41-case deterministic eval suite |
| HybridJoin Stream-Disk Engine | 🌐 Public |
Data warehouse stream-disk join engine (MySQL 8, 3-thread Python ETL) | 15,000 rec/s join throughput · Bounded hash table |
| Shortest-Path Algorithms | 🌐 Public |
C++17 graph algorithm suite (Dijkstra, Bellman-Ford, Floyd-Warshall, Johnson) | 500-run statistical benchmark with 95% confidence intervals |
| 8086 Assembly Pac-Man | 🌐 Public |
x86 MASM console game engine with ghost AI state machines & scan-code I/O | 132 KB modular ASM · Zero external runtime dependencies |
| OpenGL Snake Game Engine | 🌐 Public |
C++ OpenGL/GLUT arcade engine with collision math and state machine loop | 2,300 lines C++ · File-persisted high score engine |
| Tech-Retail-Ecosystem | 🔒 Private |
Contract-first e-commerce platform (OpenAPI 3.1, TypeScript strict, Drizzle ORM) | Zero-drift codegen · 3-tier rate limiting |
| RocketLeague-RL-Bot | 🔒 Private |
Autonomous Rocket League agent (PyTorch, Proximal Policy Optimization) | Continuous 3D flight control via reward shaping |
| Domain | Verified Stack & Tooling |
|---|---|
| AI Inference & Deep Learning | vLLM · PyTorch · TensorFlow · Mastra · HuggingFace Transformers · DFlash2 · GradNorm · AutoRound |
| Distributed Systems & Storage | Qdrant (gRPC) · Tantivy BM25 · PostgreSQL · MySQL 8 · Redis / QStash · Drizzle ORM · WireGuard |
| Languages & Systems | Python 3.11+ · C++20 / C++17 · x86 MASM Assembly · TypeScript 5 · Java 23 · SQL · R |
| Web & API Architecture | Next.js 16 (App Router) · React 19 · FastAPI · Express 5 (ESM) · OpenAPI 3.1 · Spring Boot 3.4 |
| Engineering & Tooling | Docker · CMake · Linux / POSIX / WSL2 · Git / GitHub CLI · Jupyter · LaTeX |
Crafted with engineering precision. Maintained by @ApatheticMioz.