Software and AI engineeringfor real business problems.

I build AI-enabled web applications, agent systems, computer vision pipelines and the infrastructure required to run them.

Discuss a project See selected work

Dejan Honderd outdoors in Enschede — **Dejan Honderd**

Systems built around the part that was still too slow, manual or unsafe.

Selected work across industrial classification, research agents, coding-agent workflows, streaming computer vision, speech and internal application platforms.

Work at Voortman · Industrial AI

ECLASS Classifier

From keyword search to research-backed classification.

I started with a simple TF-IDF search engine. It worked when material descriptions resembled the language in ECLASS, but struggled with abbreviations, brand names, technical shorthand and different vocabulary.

I built the benchmarking workflow on MLflow, making it fast to run engine configurations in batches and compare Top-k accuracy, rank quality and individual cases. Weighted word matching and Granite embeddings solved different parts of the problem. Reciprocal Rank Fusion produced the strongest Top-5 shortlist without requiring their scores to be comparable.

As approved classifications repeated, searching from scratch stopped making sense. Exact lookup now reuses identical accepted decisions. Nearest-neighbour history surfaces likely repeats alongside fresh search results, rather than hiding them.

Some ambiguity cannot be solved from wording alone. A manufacturer code or catalogue number may only become meaningful after finding the product online. The research agent starts from real local candidates, inspects ECLASS, searches the web and streams its progress before returning a guarded recommendation. An approver still makes the final decision. The same workflow is exposed through REST, typed SSE progress and MCP interfaces.

Discuss related work

ECLASS catalogue browser beside the live AI research activity panel while the agent investigates candidate classes — Working product · catalogue search and agent research in progress

A cleaned query first checks accepted history. An exact match fast-tracks trusted evidence; a close nearest-neighbour match is carried alongside fresh results. New wording runs through weighted word matching (TF-IDF) and Granite meaning matching, then Reciprocal Rank Fusion combines both ranked lists into search evidence. That evidence can go directly to an approver. When research is needed, the history signal and shortlist enter the agent loop, which inspects candidates, searches the web, streams its activity and returns a guarded recommendation for final approval.

Independent project · Financial research

DIT-Stack

Turn fragmented company data into a cited investment report.

DIT-Stack resolves the instrument and listing first, then normalizes filings, prices, fundamentals and macro data into provider-independent models. A coordinated research agent uses 13 domain tools, keeps an evidence workspace and assembles durable reports with citations, streamed progress and cached job state.

Identity, prices and citations are validated again before report assembly.

Explore DIT-Stack on Solaraverse

Dogfooded agent tooling

Ratchet

Make coding agents review their own work before I have to.

Ratchet is the delivery loop around the coding agents I use every day. It persists design, task and review state, then sends changes through implementation, test planning, independent review and runtime verification. Findings route back into another correction pass before a PR is ready. I keep tightening the structure as I find work I would otherwise have to catch by hand.

Eight phases. Two human checkpoints. One shared skill tree for Codex and Claude Code.

Discuss related work

Independent systems fork · Computer vision

streamSAM

A performance-tuned fork of the popular SAM segmentation foundation model, built for stable frame-ordered throughput.

streamSAM follows Meta's Segment Anything lineage through EdgeTAM. It adds bounded decode and GPU queues, backpressure, one-frame-ahead encoding and predict-then-commit memory control, so a downstream system can correct a mask before it becomes temporal memory. I use this directly in my master's thesis to provide reconstruction feedback; that application is why I built the adapter around SAM.

In the recorded 1,000-frame benchmark, the original eager path hit the 4.5 GiB RSS safety limit before processing a frame. The chunked path completed at 17.46 FPS; streamSAM reached 24.49 FPS, a 40.2% end-to-end uplift over chunked.

View source on GitHub

VoxInsert transcribing speech and inserting the text into a Windows application — Working product demo

Streaming architecture · FIFO, live inference and safe fallback

Personal open-source project · Native Windows utility

VoxInsert

Streaming dictation that is already transcribing before I stop speaking.

I started VoxInsert after thumb and wrist issues from typing so much pushed me to explore transcription. Wispr Flow and OpenWhispr were useful reference points, but I wanted a dictation tool without a browser-sized memory baseline sitting in the background. VoxInsert is a small native C++20 tray utility that inserts speech into whichever field already has focus.

Microphone audio moves through bounded FIFO and SPSC queues to OpenAI or Mistral over realtime WebSockets while I am still talking. By the time I stop, roughly 90% of a normal utterance has already been sent toward model inference. Ordered partial and final events are assembled before insertion, with file transcription as a fallback when the stream becomes untrusted.

646.5 ms median from stop press to inserted text across 10 successful Mistral realtime dictations on my development machine.

View source and download

Work at Voortman · Internal app platform

Voortman AI platform

Let employees turn process knowledge into real apps without becoming a

The Voortman AI platform lets employees use AI to turn process knowledge into small, useful applications without also taking on platform engineering. They focus on the app's behaviour and test whether it solves the work.

The platform supplies secure templates, hosting, Entra SSO, managed deployments and the complete CI/CD path through automated review, isolated testing, business acceptance and human release. Security and operations are built into the route to production instead of being recreated for every app.

The employee owns what the app should do. The platform owns the path that makes it safe to run.

Discuss related work

AI systems built beyond the model call.

For businesses with a concrete AI use case that need help turning it into a working, measurable application.

Applied AI applications

LLM and agent systems with purpose-built tools, web interfaces, cloud infrastructure, observability and evaluations against the intended use case.

Machine learning & computer vision

Inference optimization, data pipelines, evaluation and fine-tuning for teams that need more from a model and the hardware running it.

Software & platform engineering

Backend systems, APIs, integrations, dashboards and deployment infrastructure needed to make the application usable inside the business.

Lean websites

Fast, maintainable business websites and landing pages built with technical SEO and real performance constraints in mind.

About Honderd Engineering.

Honderd Engineering is an independent engineering business operated by Dejan Honderd in Enschede, Netherlands. His background is in software engineering, artificial intelligence and computer vision.

Company: Honderd Engineering
Location: Enschede, Netherlands
Registration: KvK 99802198

How I work.

UnderstandDetermine the business problem, requirements and constraints.
BuildDevelop the required software, AI system or digital solution.
IntegrateConnect it to existing systems and workflows where necessary.
SupportImprove, maintain or extend the system after deployment when required.

Have a software, AI, automation or web project?

Send the problem, the current situation and what a useful outcome would look like.

Email directly[email protected]