How Apuna works — and what AI actually is.

Whitepaper — Open by Default, Reliable by Subscription

Our first paper, co-authored by JTO and an Apuna AI agent (disclosed as such): how we build AI you can audit — a documented loop where agents propose, a review panel scores, a verification step checks every claim against the rendered page, and a human greenlights each change before it ships. The proof is reflexive: this site was built that way. Read it on the page, or download it as a PDF.

How an engagement works

Every engagement starts the same way: you bring your brief and, if you have one, your company URL. We run discovery and deep research to understand the operation, the constraints, and the decision you actually need to make. From there the crew builds in a stream of small, contained changes — each one reviewable before it ships. When the work is done, the application runs on your infrastructure, the code is licensed Apache-2.0, and it is yours. No proprietary cage, no lock-in clause. The expensive part of leaving would be replacing the people who know your system — and most clients find that is a reason to stay, not to leave.

Read the page — not the code

This is the most important thing to understand about how we build. Every change we make is rendered for you as a real page before it ships. You review what the change looks like and does — not the underlying code. You greenlight it, it ships. You push back, we revise. The decision is always yours, made on the thing a real user would see, not on a wall of coloured code. That principle is not an aspiration — it is a structural constraint we build every engagement around.

Apuna Care & support

After handover, Apuna Care is the subscription that keeps your application running. Three tiers: Community is self-service — the open-source documentation and best-effort community support, right for teams who want to run it themselves. Standard adds managed maintenance: patches, security updates, and business-hours support with defined response times. Premium is full assurance: everything in Standard, plus priority incident response across extended European hours (currently around GMT±1), a target of approximately one hour response on critical issues within those hours, proactive monitoring, and a named contact who knows your system. We commit to response time — when a specialist starts work — not a guaranteed fix clock, because real systems vary. No prices on this page; talk to an engineer. Follow-the-sun coverage across more time zones is the direction we are heading — one zone at a time as the team grows.

Commodity Intelligence

Commodity Intelligence is a live tool available at /commodity. Paste your company URL: the system surfaces up to five key purchasing commodities — the materials or inputs that drive the most value in your direct procurement, weighted accordingly. It then tracks spot and futures markets for those commodities in a live dashboard. The purpose is to give your procurement team evidence-based timing for buying and hedging decisions. It is decision support, not financial advice, and the judgement call is always yours.

Your data & privacy

The AI in our tools is disclosed, never disguised. The full privacy policy — what we collect, why, the legal basis, your rights, and the sub-processors involved — is at /datenschutz. If you have a specific question about how your data is handled in an engagement, write to hello@apuna.dev. A human answers.

Getting started

If you are ready to scope a project or just want to understand what is feasible for your operation, the contact form is the right first move. Tell us where you are today and where you want to be. We respond with a concrete plan — not a sales deck. Typical response: within one business day.

Hardware Bring-Up — grounded answers for electronics work

Name a component — a microcontroller, a sensor, a power regulator — and optionally give us a product page URL. The tool fetches the manufacturer's documentation, resolves the exact variant you have, and gives you wiring diagrams, configuration steps, and sample code grounded in what the datasheet actually says. Where the documentation is silent, it says so explicitly rather than guessing. This is a free experimental tool; verify every claim against the manufacturer's datasheet before you connect anything. It runs on free models and is publicly available at /hardware-bringup.

Office Plant Care — the open core on itself

A weather- and season-aware watering log for our own office plants. It sounds trivial — that is the point. A well-scoped AI tool solves exactly one decision (when does this plant need water?) with real data (today's weather, the season, the last watering date) and nothing it does not need. It is open-source, public to read, and exactly the kind of single-decision application we build for clients, shown here on ourselves. The code is the documentation.

Getting value from the open-source core

Everything Apuna builds runs on apuna/core — an open foundation licensed Apache-2.0. You can take the core, run it on your own servers, and build from there without asking permission or paying a licence fee. What you get for free: the code, the architecture, the agent scaffolding, and the documentation. What Apuna adds on top: the people who understand your business, the engagement process that scopes the right problem, and the Apuna Care subscription that keeps it running. Open source does not mean unsupported — it means you are never locked into us for the code itself. The three Labs tools (Commodity Intelligence, Hardware Bring-Up, Office Plant Care) are the first examples of what gets built on it.

Start from scratch — what AI actually is

Think of an AI model as a very well-read assistant who has absorbed an enormous amount of text — books, articles, code, conversations — and learned patterns from all of it. Ask it a question, and it predicts, word by word, what a useful answer looks like. That is all it does. It is not thinking the way you think. It has no memory of you between conversations. It does not know what is happening in the world right now unless someone tells it. What it is genuinely good at: drafting, summarising, translating, finding patterns in large amounts of text, and generating first versions of things that a human then improves. What it is bad at: knowing when it is wrong, keeping facts reliably accurate, and making judgement calls that require knowing your business. The first practical step for any manager: pick one real task you do every week that involves reading or writing, try it with a simple AI tool for two weeks, and notice what it gets right and what it gets wrong. That small experiment will teach you more than any amount of reading about AI.

The words, in plain language

AI agent

a program that can take actions on your behalf, like sending an email or searching the web, not just answer questions

Model

the trained system that generates the text; what most people mean when they say "the AI." Token — the small chunks a model reads and writes; roughly three quarters of a word on average, which is why AI bills are counted in tokens

Prompt

the instruction or question you give the model; better prompts get better results, every time

API

a connection point that lets two software systems talk to each other; when Apuna connects an AI model to your business system, an API is usually the bridge

Open source / Apache-2.0

code you can read, copy, modify, and use commercially without asking permission; all Apuna-built software ships under this licence

LLM

Large Language Model; just another name for the kind of AI that reads and writes text

Human in the loop

a design choice where a person reviews or approves what the AI proposes before it acts; Apuna builds this in by default

Cloud vs on your own computer

running AI on a provider's servers means lower setup cost but your data leaves your building; running it on your own hardware keeps data local but costs more upfront

As much AI as needed, as little as possible

The goal is not to use more AI. The goal is to solve the problem. Example one: a company receives the same ten types of supplier email every week. A simple routing rule — no AI — sorts nine of them automatically. Only the tenth, the ambiguous one, goes to an AI model to classify. Cost: a fraction of running everything through AI, and the nine easy cases are never wrong because they never touch a model. Example two: a purchasing report that used to take a person two hours now gets its data pulled and formatted automatically by a script. An AI model reviews the final version for anomalies and flags anything unusual. The AI does ten minutes of work; the script does the ninety. Example three: a customer FAQ. Ninety percent of questions are answered by a static page with a search box — fast, reliable, no AI cost. The AI handles only the questions the static page cannot answer. The principle behind all three: identify the part of the task that is actually hard and variable, and apply AI only there. Everywhere else, a rule, a template, or a script is cheaper, faster, and easier to check.

What "AI integration" actually means

Imagine your team's quoting process. Someone reads a customer email, opens a spreadsheet, copies product codes, applies the right margin, and drafts the quote. That takes forty-five minutes per quote. AI integration means adding a step where a model reads the customer's email, pulls the relevant product codes from a database, and drafts a quote for a human to review and send. The human still checks and approves — they just spend five minutes rather than forty-five. That is integration: the AI handles the mechanical reading and assembly; a person handles the judgement and the send. The useful question is always: which part of this task is mechanical and repeatable? That is the part worth integrating. The rest stays with a person.

What "automation" actually means

Automation is older and simpler than AI. A spreadsheet formula is automation. A scheduled script that runs a report every Monday morning is automation. What AI adds is the ability to handle inputs that vary — text that comes in different shapes, questions phrased differently each time, documents that are not all in the same format. Before reaching for AI, Apuna always asks: can this be solved with a rule? Rules are deterministic — they always do the same thing given the same input, which makes them easy to test and trust. AI is a better fit when the input genuinely varies and a rule would need to cover too many cases to be maintainable. A well-designed system uses both: rules for the predictable, AI for the variable.

What "process analysis" actually means

A production manager who has been on the floor for fifteen years makes certain decisions by instinct. She knows which supplier tends to slip on Fridays, which machine needs an extra check when humidity climbs, and which customer is worth expediting a rush order for. That knowledge lives in her head. Process analysis means making that knowledge visible — capturing it in a dashboard or an alert system so it can be acted on even when she is not there, and so it does not walk out the door when she retires. It is not about replacing her judgement. It is about recording it, scaling it, and giving the next person a foundation to build on. The first step is almost always the same: find out what data already exists in the operation, and make it readable before doing anything else.

More words, in plain language

RAG

Retrieval-Augmented Generation; a technique that gives an AI model access to a specific set of documents at query time, so it can answer questions about your own data rather than only what it was trained on

Fine-tuning

adjusting an existing model on new examples so it responds in a particular style or domain; rarely necessary and often over-sold as a solution

Hallucination

when a model produces a confident-sounding answer that is simply wrong; the main reason a human must review AI output in any high-stakes context

Guardrail

a check or constraint that prevents the model from producing certain outputs; can be built into the prompt, the surrounding code, or both

On-premises (on-prem)

software that runs on hardware inside your own building, under your control; Apuna builds this way by default

Open-source

software whose source code is publicly available; Apache-2.0 is the licence Apuna ships under, which allows commercial use without restriction

Inference

the act of running a model to generate an output; what happens when you send a message and the AI responds

Context window

the maximum amount of text a model can read and hold in attention at once; once you exceed it, earlier content falls out

Embedding

a mathematical representation of text that lets a system find semantically similar passages without exact keyword matching; the mechanism underneath most search-and-retrieve features

Recipe: sort your emails with a rule before touching AI

This recipe has no AI in it — and that is the point. Step one: for one week, keep a tally of the emails that land in your team's shared inbox. Group them by type — order confirmations, supplier delays, complaints, internal requests, spam. Step two: count which types appear most often. Step three: in your email client (Outlook, Gmail, or almost anything else), create a filter rule for the most common type. Example: if the subject line contains "Order confirmation" and the sender is in your supplier list, move it to a folder called "Confirmed orders" and mark it read. Step four: wait one week and measure how much time the team saves on manual sorting. Most teams find that 60-70% of their inbox volume falls into three or four predictable types. Rules handle all of those — for free, instantly, and without ever being wrong. AI is worth considering only for what is left.

Recipe: summarise a long document in two minutes

This one uses an AI tool — and it takes about two minutes once you have done it once. What you need: access to any AI assistant (Claude, ChatGPT, or similar; most have a free tier). Step one: open the document you want to summarise. Select all and copy. Step two: open the AI assistant and paste this prompt exactly, then paste your document underneath it: "Summarise this document in five bullet points. Each bullet point should be one sentence. Do not add anything that is not in the document." Step three: read the output. Check each bullet point against the document. The model will almost always be accurate on a factual document — but always check, especially for numbers, names, and dates. Step four: if a bullet point is wrong, paste it back and say: "This point is wrong. The document says: [paste the correct passage]. Please correct it." That is a complete workflow. The two guards: always check the output, and never paste confidential documents into a cloud AI tool without understanding your organisation's data policy.

Recipe: classify a list of items with a prompt

Useful when you have a list of things — customer complaints, supplier notes, product descriptions — and want to sort them into categories without reading every one. Step one: decide your categories in advance. Write them down. Example: "Delivery issue", "Quality issue", "Billing issue", "Other." Step two: paste this into your AI assistant, filling in the category names and the list: "Classify each item in the list below into one of these categories: [your categories]. For each item, respond with only the item number and the category name. Do not explain. List: [your list here]." Step three: the model returns a numbered list of classifications. Step four: spot-check ten items by reading the original and comparing. If the error rate is low enough for your purpose, the classification is done. If not, add more specific instructions to your prompt: "If a complaint mentions a missing parcel, classify it as 'Delivery issue.'" Adjust until the accuracy is right. This works reliably for lists of up to a few hundred items in a single session.