I design practical systems powered by AI
I’m James Dye, founder of XyFlow Innovations. I build cross-domain tools — from GNSS timing analysis and grid simulations to automation workflows, arbitrage engines, and AI-assisted creative systems.
XyFlow is my independent R&D space for fast iteration, applied experimentation, and building real prototypes from complex ideas.
My work focuses on practical, high-leverage systems — built quickly, validated with data, and refined through iterative development.
Selected R&D Work
I build systems across multiple domains — spanning water treatment, satellite timing analysis, automation workflows, and AI-augmented tools. Here are a few examples.
AquaSol V10 — Theoretical Electrodialysis Desalination Architecture
AquaSol V10 is a theoretical desalination system design built around a novel dual-stack toggle electrodialysis (ED) configuration. The architecture alternates two ED stacks between heavy-duty and light-duty states, enabling self-flush scaling mitigation while maintaining continuous operation. The goal of the design is to reduce fouling, improve runtime stability, and lower operating cost for small, off-grid desalination units.
The system remains conceptual and requires full laboratory validation.
However, the design advanced through the evaluation stages of the XPRIZE Water Scarcity | Mohamed bin Zayed Water Initiative, earning ~$36,000 in milestone award funding for reaching the Qualified Teams Testing round.
Status: Concept-stage design; pending experimental validation.
Documentation: High-level system brief available upon request.
Competition: XPRIZE Water Scarcity — Qualified Teams Testing (Milestone Award Recipient).
GNSS Error Detection & Predictive Timing Model
Multi-Year Satellite Data Analysis for Predicting 3–5 ns Severity Events
This project analyzes nine years of IGS Final, Rapid, and Ultra-Rapid GNSS clock and orbit products to detect and potentially predict small-scale satellite timing anomalies. The system uses custom “lever” features spanning satellite geometry, atmospheric conditions, clock type, and historical error behavior.
The resulting model identifies 3–5 nanosecond error episodes — including some events that Rapid solutions miss — with low false-positive rates. The goal is to provide surveyors, drone pilots, and precision-GNSS applications with the ability to mask specific bad epoch windows during jobs.
Status: Production-bound. Web interface in development (launching soon).
Tools: Python, NumPy, SciPy, custom feature engineering, AWS cloud-hosted VM pipelines, and iterative model tuning assisted by ChatGPT/Codex.
AI-Assisted Long-Form Writing & Canon Engine
An iterative, writer-in-the-loop system for structured long-form creation
After roughly a month of weekend experimentation, this project evolved from an attempt at a fully autonomous book generator into a practical writer-in-the-loop engine. The system supports long-form storytelling by maintaining structured memory objects — character sheets, locations, timelines, canon notes, and evolving story facts — enabling chapter-by-chapter drafting with continuity checks.
The engine blends local LLaMA-13B and Mistral models with a set of specialized agentic API calls, each performing different editing and reasoning roles. A custom gating logic layer arbitrates these agents, and an internal metric extractor evaluates tone, pacing, and structure using features derived from published books. Together, these components keep drafts consistent while supporting iterative rewrites without exposing the underlying proprietary workflow.
The result is a controllable, refinement-focused tool: not “fully autonomous,” but a reliable assistant for authors who want structured creativity without losing coherence.
Status: Active development. Core memory system and rewrite loop functional; export tools and UI planned for next iteration.
Tools: Local LLaMA-13B, Mistral, Python scripts, multi-agent prompting workflows, gating logic, and style-metric extraction modules.
HaloLite v1 — Spherical Multi-Band EV Concept
Concept-level vehicle study of a rolling spherical “halo” pod with a gimballed seat, layered crash structure, and human / drone operating modes.
HaloLite v1 explores what a ground vehicle could do if the occupant sat inside a protected spherical shell instead of a conventional car body. The outer “halo” rolls on multiple treaded bands, while an inner pod and seat gimbal keep the rider aligned and tightly restrained. In normal “human mode,” the system targets small-EV behavior: brisk straight-line performance, controlled g-loads, and everyday energy use in the same general band as a compact electric car.
Using a Python-based modeling sandbox (analytic dynamics, crash approximations, and multibody visualizations), I compared the halo layout against simple rigid shells and monowheel-style baselines. Within those shared assumptions, the layered halo structure shows directional benefits in crash softening, and the multi-band geometry enables evasive “juke” maneuvers and spin-brake behaviors that are hard to match with a conventional layout.
Status: Concept-stage engineering study; no hardware built.
Stack: Python modeling sandbox (NumPy, analytic dynamics, simplified crash models, multibody visualizations).