BI Documentation That Actually Gets Read

A team with dozens of PowerBI semantic models had no documentation. New team members spent weeks reverse-engineering measures and relationships.

We built a pipeline that extracts complete metadata, generates descriptions using GPT-4, maps relationships via Graphviz, and publishes to Markdown, HTML, JSON, or Wiki format.

Outcome: The team cut onboarding time from weeks to days. Documentation stays current because the pipeline regenerates when models change.

Data Intelligence — Python, PowerBI, GPT-4

Autonomous Workflows Without a Middleman

A client needed to automate multi-step technical workflows: fetch data, analyze it, generate code, deploy changes, and report results. Separate tools and human handoffs were slow and error-prone.

We deployed a multi-agent orchestrator on the Hermes Agent framework that decomposes complex tasks, delegates to specialist agents, executes code, and synthesizes results.

Outcome: The client reduced a 4-hour manual process to 15 minutes of review time.

AI Automation — Python, Hermes Agent, MCP

A 3D Racing Engine, in a Browser, No Install

A game studio needed a Wipeout-style racing game playable in a browser with no installation. They needed procedural tracks, AI opponents that raced fairly, and physics that felt authentic.

We built a complete WebGL engine in Three.js with spring-damper hover physics, AI at 5 difficulty tiers, procedural track generation, weapon systems, and a Forza-style scoring engine.

Outcome: The studio had a playable prototype in weeks. Every AI race is a fair race.

Game Engineering — Three.js, WebGL, AI

482 Spells, One Campaign, No Manual Data Entry

A D&D 5e campaign setting had custom subsystems, knowledge-gated content, and over 400 spells that needed to work inside Foundry VTT. Manual entry would have taken months.

We developed a module ecosystem with spell system overlays, cognitive load tracking, 3D canvas replacement, encounter automation, and a Plutonium import pipeline converting 482 spells automatically.

Outcome: The campaign went live with all content available day one.

Foundry VTT — TypeScript, Three.js, Content Systems

Ten Services, One Server, No Downtime

A client needed to run Matrix, OnlyOffice, Foundry VTT, Home Assistant, and multiple AI services on a single self-hosted stack. Each had different networking and security requirements.

We designed a Docker Compose platform with reverse proxy, CSP headers, nightly backups, and health monitoring. Every service has a documented runbook.

Outcome: Months of uptime without unplanned downtime. New services are a Docker Compose entry and a runbook away.

DevOps — Docker, Nginx, Linux

Meeting Transcription That Knows Who Said What

An organization needed accurate meeting transcriptions with speaker identification. Off-the-shelf solutions failed on technical vocabulary and speaker separation.

We deployed NVIDIA NeMo ASR with speaker diarization, custom audio preprocessing, and a web frontend for reviewing results. Output is timestamped and speaker-labeled.

Outcome: The organization now searches past meetings by speaker and topic.

Speech — Python, NVIDIA NeMo, ASR

An AI Terminal That Fits in Your Hand

A concept needed a portable device running AI chat conversations with a touchscreen interface, WebSocket connectivity, and swappable AI backends.

We architected a system using the CYD ESP32 display with LVGL UI, a Python WebSocket server, and a pluggable agent runtime supporting Direct LLM and Hermes bridge.

Outcome: Whiteboard to working prototype with a documented build spec.

Embedded — ESP32, Python, WebSocket

DataFlow ETL Framework

Published on PyPI as etl-framework v1.0.1. A fully JSON-driven ETL pipeline framework with enterprise-grade security, RBAC, encryption, and audit logging.

Outcome: New pipelines are deployed by writing a JSON config. The framework handles extraction, validation, transformation, and loading.

Data Engineering — Python, pandas, PostgreSQL, PyPI

Universal BI Platform

A full BI platform using Cube.js, React, and PostgreSQL with multi-tenant isolation, real-time analytics, and a visual query builder for non-technical users.

Outcome: Replaces proprietary BI tools, eliminating per-seat licensing costs.

BI Infrastructure — Cube.js, React, PostgreSQL, Docker

Local RAG Pipeline

A fully local RAG system with document ingestion, embedding generation, vector search, and LLM-based answer synthesis. Everything runs on-premises with no external API calls.

Outcome: Organizations query their documents using natural language without sending data to third-party services.

AI Infrastructure — Python, Vector Database, Docker, Local LLM

Open WebUI Theme Engine

A complete theming system for Open WebUI that generates CSS from JSON theme definitions. Supports component-specific styling, gradients, glows, and unlimited themes.

Outcome: Organizations deploy branded AI chat interfaces matching their design systems.

AI UI — CSS, JavaScript, JSON, Open WebUI

Character AI Bot

A hexagonal-architecture AI personality bot with agentic memory, modular personality definitions, and persistent conversation state.

Outcome: Multiple personality profiles deploy, swap, and extend independently.

AI Architecture — Python, FastAPI, Redis, Hexagonal Architecture

Docker Run to Compose Converter

A full-stack SvelteKit web app that converts docker run commands into production-ready Docker Compose YAML. Three-tier UI, security validation, examples library, and comprehensive documentation.

Live site: dockerhelper.npc-it.co.uk

Full-Stack Web — SvelteKit, Express.js, TypeScript, Docker Compose