Skip to content
AssessmentDERU

AI Software Engineering

AI works fast. I work precisely.

With AI, I build you the tool for the task that comes back every week, in a few days to a few weeks.

Request a free initial assessmentReply within 3 business days
The tool is new. The craft is not.

Systems I helped build have run in automotive and medical-device manufacturing, at international live sports events, and in camera networks on three continents.

Does this sound familiar?

Six situations I keep encountering in small and medium-sized businesses.

ENGILO in four answers

01Why now?
Agentic AI development means: I describe what should exist in the end, an AI agent builds it, and I check every result. This has drastically lowered the cost of custom software. What used to take an agency budget is now within reach for any SMB, especially where you earn your money: the quote, the order, the customer question.
02What makes it special?
Twenty years of real-time and AI systems have shown me what matters when software has to run every day. That knowledge goes into every tool I build today.
03What if I've been let down before?
Maybe you've already paid for software that looked good in the presentation and then didn't hold up in daily use. After that, you tell yourself: never again. That's why, with me, what has to work at the end is settled before the project starts, and you pay a fixed price for exactly that result.
04What happens next?
Tell me about one process that costs you time or revenue every week.

Request a free initial assessment

Case studies

Real projects first, prototypes and own developments after. Every case carries its label.

01An AI assistant that answers from the company's own documentsClient projectShow details +
Starting point
An online community received the same questions from its users every day. The answers had long been in their own documents, but nobody found them quickly enough there. Support work ate up time every single day.
Task
An assistant that answers from the company's own knowledge rather than from general internet knowledge. The document collection was to stay in house, and an employee had to be able to take over at any time once a conversation got personal.
Solution
An AI assistant on a server belonging to the client. It finds the relevant passages in their own documents and phrases the answer from them instead of guessing. The documents and their search database stay on that server; only the matching excerpt is sent to an external language model to phrase the answer. Anyone already running their own model can connect that instead. Employees see every conversation and can take over at any point. Connections already available: Telegram, team chat and the community app, more are possible. Data protection built in from the start, including user consent.
Result
In production since August 2026. First client live, final polish under way.
Stack
Python · FastAPI · PostgreSQL + pgvector · Ollama (optional) · Docker
  • 3 connections, more possible
  • documents on the client's own server
  • external or self-run language model
02Legacy software secured, during live operationClient projectShow details +
Starting point
Two connected applications from 2010, in daily use for years, but with security at yesterday's standard: vulnerable password storage, credentials in the source code, and no reliable way to ship a new version.
Task
Secure them without interrupting daily operations, and give an honest number for what a thorough rebuild would cost.
Solution
First a complete inventory: 22 vulnerabilities, sorted by urgency. Then the critical ones fixed: secure password storage (migrated invisibly for users), protection against automated attacks, credentials removed from the source code. Plus an automated safety net of tests and a deployment that is reproducible at the push of a button.
Result
Both applications have been running secured with the real production data since August 2026. The further rebuild is priced, and the client decides at their own pace.
Stack
Java EE · MariaDB · Docker · GitLab CI/CD

Project ongoing; only completed work is mentioned.

  • 22 findings fixed
  • password protection modernised
  • 644 tests, 86% coverage
  • push-button deployment
03Website relaunch in four languages for an international wealth advisoryClient projectShow details +
Starting point
The old website came from a site builder, existed only in English, and loaded its images from third-party servers. For an international wealth advisory that was neither professional nor privacy-friendly.
Task
A multilingual website that actively wins clients: prospects request a whitepaper and land automatically in the sales system.
Solution
New website in English, German, Spanish and Chinese, with all content and media on their own environment and legal texts in all four languages. Whoever requests the whitepaper briefly confirms their email address and is automatically added to the sales team's CRM.
Result
The site is live. New prospects flow into the sales system without any manual work.
Stack
Next.js · next-intl · Docker · Traefik
  • 4 languages
  • 12+ pages per language
  • 0 third-party CDN calls
04One platform, several sitesClient projectShow details +
Starting point
A group of companies needed several web presences at once: a news and insights portal with an editorial area for a magazine, plus an information portal for an investment product that keeps all due-diligence material in one place. Building each site separately would have meant doing the same work three times.
Task
A shared foundation from which every site can be derived in days, instead of starting from scratch each time.
Solution
A master platform that is derived per brand: everything brand and project specific lives in a single configuration file, the rest stays shared. It includes an editorial area with a rich-text editor, roles for editors and administrators, categories, an automatic RSS feed and imports from spreadsheets and text files. The information portal for the investment product builds on the same foundation: one page with 19 tabs from terms and collateral to key figures, documents readable directly in the browser, per-table export, full-text search across all data fields, and news drawn from the shared editorial database. It runs either in the cloud or as a container on the client's own infrastructure.
Result
Both sites are built and are currently being finished. The second one came out of the same foundation as the first.
Stack
React · TypeScript · PostgreSQL · Edge Functions · Docker

Projects ongoing; only finished work is mentioned.

  • 1 foundation, several brands
  • 19 tabs on one page
  • editing, RSS and import built in
05Customer portal for a beverage wholesalerPrototypeShow details +
Starting point
A beverage wholesaler took orders by phone, messenger and email, and everything was transferred by hand. On top of that, every bar or restaurant has its own assortment with its own prices.
Task
A quote for a website plus ordering portal. To make it tangible, the prototype was built as part of the quote itself.
Solution
A working test version: every business logs in and orders from exactly its own assortment. The wholesaler sees all orders in one place, with status and history. Articles are imported straight from the ERP system, so nothing is maintained twice.
Result
Shows what two weeks can produce. The quote was not commissioned. The prototype now serves as a demo with sample data, not as a reference.
Stack
React · TypeScript · PostgreSQL

Prototype from a quote that was not commissioned.

  • 193 articles
  • 150 assortments
  • 175 businesses
06How I prove qualityOwn developmentShow details +
Starting point
An own development as a stress test: a simulation core in which saving, replay and multiplayer only work if every machine computes exactly the same result. The smallest deviation is a bug.
Task
Prove that the software stays exactly reproducible while it keeps growing, with tests instead of promises.
Solution
Every run is recorded and can be replayed step by step; checksums compare every intermediate state automatically. On top of that, dedicated tests watch the source code itself: they trip as soon as anyone breaks the agreed rules, an AI included.
Result
Over 1,300 automated tests green; a complete run is repeated bit for bit. It answers the question everyone has about AI-generated code: how do I know it's right?
Stack
C# · .NET · Godot
  • 1,300+ tests
  • 8,859 steps repeated bit-perfect
  • 4 source-code guards

Services

After the free initial assessment you receive a fixed-price quote. Code and documentation belong to you, and the system runs in the environment of your choice, which you can change at any time.

  • PROZESS-AUTOMATISIERUNG▶ Entry point

    Process automation

    Orders from email and messenger land directly in your inventory system, and the quote is drafted from the meeting notes. Built as a maintained system with tests, documentation and hand-over.

    Price
    from €3,500, fixed price after initial assessment
    Duration
    2–3 weeks
  • KUNDENPORTAL

    Customer portal

    Your customers order, book and check their status themselves. Admin area for your team.

    Price
    from €6,000, fixed price after initial assessment
    Duration
    4–6 weeks
  • KI-ASSISTENT▶ Entry point

    AI assistant on your data

    Answers from your own documents (RAG). An external language model phrases the answer; your own can be connected instead. Users know they are talking to an AI, and a staff member can take over at any time.

    Price
    Hosted subscription (cancellable monthly) or on-premises installation from €4,900 at a fixed price
    Duration
    4–8 weeks

    Client login: rag.engilo.de

    Your documents and data stay yours. The assistant software is licensed.

  • WEBSITE

    Website with its own backend

    Multilingual, fast, with content, orders and a customer area, and no site-builder subscription. Legal texts prepared, final review by your lawyer.

    Price
    from €3,000, fixed price after initial assessment
    Duration
    3–4 weeks
  • BEGLEITUNG

    Technical guidance

    Securing legacy systems, architecture advice, code review, second opinion.

    Price
    from €700, fixed price per engagement
    Duration
    as needed
  • BETREUUNG

    Operations & support

    Server, updates, backups, monitoring, business-day response times, named substitute. Available with every package, cancellable monthly.

    Price
    from €150 per month, cancellable monthly
    Duration
    ongoing

All prices exclude statutory VAT. The starting price is the lower bound; the initial assessment names the figure for your project.

Twelve questions you should ask any AI provider

„One person: what if you're unavailable?“
Code and documentation are yours from day one, and hand-over readiness is an acceptance criterion. For operations there is a support agreement with response times and a named substitute.
„AI code is throwaway code.“
Throwaway code is what you get when nobody looks. Here the agent delivers, and I check whether the result is good or merely finished, with automated tests that prove it. The responsibility stays with me.
„Do my data end up with a US corporation?“
Your data and documents stay in the environment you choose. With the AI assistant, only the matching text excerpt ever leaves the building, and even that only if you don't run your own language model.
„What is the AI doing that I can't see?“
An agent gets tighter limits from me than any employee: it reads only what it is meant to read, every change passes automated tests, and whenever it deviates from the specification, you get a question. Text from emails, documents or websites is treated as data; instructions hidden in there are not carried out.
„We're too small for AI.“
The initial assessment gives you an honest answer on whether it's worth it. Sometimes the answer is no.
„What does it cost?“
A fixed price you know after the initial assessment and before the project starts. The starting price on each card is the lower bound. The one exception is the AI assistant, a ready-made solution as a subscription or installed on your premises.
„What about backups?“
Backups are part of the operating environment I set up. What is backed up, how often, to which separate storage and how the restore is verified is written into the specification of every project before I build.
„Who is liable if something breaks?“
Acceptance is against an acceptance protocol with testable criteria; if the system deviates from it, I fix it. Beyond that, statutory liability applies, with its scope defined in the contract.
„We run on Microsoft 365 and an inventory system, and some things must not leave the building.“
Existing systems stay. What I build connects through interfaces or imports. Whether your case can run entirely on your premises, and what a local language model delivers then, I tell you in the initial assessment after reviewing your environment.
„What if the AI loops overnight and the bill explodes?“
Two limits: a hard spending cap in the language-model provider's account, at which calls stop, and per-user and per-day ceilings inside the software itself, with an alert when they are exceeded.
„Will my team still be using this in four weeks?“
Hand-over includes training your team and onboarding a named contact person in your business. Four weeks after launch we check together: what is being used, what is getting in the way, what needs improving.
„Next year a better model comes out. Do I start over?“
No. Your knowledge, your workflows and your rules live in files and in the specification; the language model is a replaceable component. A new model gets connected, everything else stays.

How I work

I build software for a business in four steps. Each step has a defined result that you can see and check before the next one begins.

  1. 01

    Understand

    We talk about the process, not about software: who does what today, with which data, where it gets stuck, what should measurably improve. I keep asking until I could explain the workflow without your help, and we agree on what should not be built. Along the way we write down what so far lived only in people's heads: rules, exceptions, the way your customers talk.

    What you get

    A written problem statement of one to two pages: starting point, goal, users, success criterion, scope boundaries. You confirm or correct it.

  2. 02

    Specify

    The problem statement becomes a specification package: user roles, testable acceptance criteria per feature, edge cases, data model, interfaces, operating environment. I make the technical decisions here and justify them.

    What you get

    The specification package as readable documents, a fixed-price quote, and a project skeleton with version control and automated checks.

  3. 03

    Build: agentic and test-secured

    Implementation feature by feature with AI coding agents, each in three phases: plan, implement, verify (automated tests, type checking, browser testing). Only when all checks are green is the state versioned. Critical parts are cross-read by a second, independent model. The AI never works without a specification, never without tests, never on your production systems.

    What you get

    After every feature, a running version on a test environment you can try yourself, plus a change log.

  4. 04

    Hand over and operate

    Delivery into the environment of your choice (containers, reverse proxy, TLS, backups): your own infrastructure, a hosting provider you trust, or an instance I look after for you. Credentials stay with you, never in the code. I hand over source code, specification, operations manual and acceptance protocol, train your team, and four weeks after launch we check together what is being used and what is getting in the way.

    What you get

    A system that belongs to you and that any other provider could take over, without me and without being tied to one platform.

Portrait of Stefan Diener

About me

Stefan Diener, Dipl.-Ing. (FH)

Precision mechanic, engineer in optical technologies and image processing, twenty years of real-time and AI systems in employment, independent since April 2026.

  • Co-inventor of 2 patents (video matching in camera networks, measuring audience attention)
  • Talks: AWS re:Invent 2017, WKF 2022 (Korea)
  • Based in Rhein-Main / Südhessen, Germany, with on-site meetings within the region
My last name, Diener, is German for servant. In Esperanto it becomes servilo, the word for server. I didn't plan that, but it fits: I build systems that serve others, and I make sure they keep running.

Career

  1. 1990s

    Precision mechanic

    Precision as a craft.

  2. 2000s

    Dipl.-Ing. (FH) in optical technologies and image processing

  3. 2002–2011

    Industrial image processing

    For automotive and medical-device manufacturers.

  4. 2012–2026

    Senior engineer at an international AI company

    Including 2015–2019 event and sports tracking for international martial-arts, basketball and fashion events; Coach AI SDK, 3D pose.

  5. 2026

    Independent, ENGILO

    Engineering firm for software & AI. Agentic Coding Master Class.

Contact

Leave to me what costs you time every week. Write to me, I'll take care of it. Within 3 business days you'll receive a written assessment: what's worth doing, what isn't, and roughly what it costs.

I use your details solely to answer your inquiry (Art. 6(1)(b) GDPR) and delete them no later than 6 months after our last contact if no engagement results. Details in the privacy policy.