EthosPowerAIDelivery
Steps 3 and 4 of the plan: build on the data your clients' AI cannot reach, prototype with them in weeks, then run it as an auditable service with a person signing off each decision. Open-source, on your infrastructure, yours to keep. Every build makes a named person better informed; none replaces one.
Automations · Step 4
Stop doing it by hand
Work that moves by hand between systems that don't talk: test reports into the asset register, approvals and handoffs, notifications, bridges to legacy systems.
What this looks like in practice
A field engineer's battery discharge-test reports — voltage, temperature, and discharge current across up to 54 battery blocks — captured by voice through a privacy-first AI agent running on the company's own server.
AI Agents · Step 3
An assistant that knows your records
Assistants that research, draft, triage and answer from your own records rather than the public web, for support, sales, research or multi-step workflows. Built on LangChain/LangGraph with Claude, OpenAI or a local Ollama model, and Qdrant or ChromaDB for retrieval.
What this looks like in practice
An energy-sector company was losing leads to manual follow-up. I built a capture system that checks the ERP/CRM for a match on email, company or phone before scoring a lead: duplicates merge into the original record, and new contacts get a three-email welcome sequence behind a person's approval.
Websites & Web-Apps · Step 4
Your public presence, or a tool your team logs into
Websites that answer an engineer's question from your own documents, and tools your staff and clients log into: dashboards, portals, intake and approval apps. This site runs on the same stack, Next.js 15, Payload CMS and Cloudflare Workers, and you own the code.
What this looks like in practice
ebatt.ai grew from spreadsheets, through Django web apps, into a live battery-quoting engine: IEEE 485 sizing and EN 50272-2 venting free, selection, RFQ and checkout behind them. One developer built it, which is why your first tool can start small too.
MCP Connectors · Step 3
Make your systems answerable from your customers' own AI tools
A prospect asks their AI before they ask you. A Model Context Protocol connector lets that AI query your systems and get your answer rather than a guess from whatever it scraped, under the access rules and audit trail you set.
This site runs one at /api/mcp, serving its engineering knowledge base; point your own assistant at it and see how the pattern behaves before you commission yours.
Platforms · Steps 3 and 4
The whole system, end-to-end
The battery example on the approach page as one system: charger test results in, a replacement forecast, a check on stock and lead times, the site work scheduled, and a person approving the purchase order. Built on your infrastructure as one system, not five point tools. Discovery first.
What this looks like in practice
An engineering consultancy wanted its people to search decades of reports and design documents with AI, without exposing them to cloud models or anyone outside the firm. I built a self-hosted knowledge base with document management and end-to-end encryption; the archive never leaves the firm's servers.
All prices in USD (converted from NZD), exclusive of GST. Advisory rate: $125/hr.
Not sure which of these fits?
Start with a call, not a quote. If a small automation solves it, I'll say so before I pitch you a platform.