II. Works
The Gallery
Client work is unnamed by agreement
I
Forecasting · Full stack
Built for a working restaurant. Three decompositions find the patterns in a year of orders, an LSTM projects the fortnight ahead, and it lands as a shopping list.
II
Conversational commerce · Agents
An ordering and customer service agent living inside WhatsApp. An intent router splits orders from purchases, and nothing reaches a customer unapproved.
III
Streaming · Distributed ML
Sentiment on Amazon food reviews as they arrive, not in a nightly batch: Kafka streams, Spark MLlib classifies, MongoDB keeps. Team of four; I built the pipeline.
IV
Computer vision · Deep learning
Upload a video and get a verdict with its reasoning attached. It samples sixteen usable frames, scores each with a fine-tuned EfficientNet-B0, and shows the numbers.
V
Sales intelligence · Research
Type a Dutch company name and get the briefing a salesperson would spend an afternoon assembling: register data, web signals, pain points, a one-page PDF.
VI
Enterprise assistant · RAG
An assistant inside Microsoft Teams that answers staff HR questions in Dutch and files IT tickets without leaving the chat. Every answer cites its source.
VII
Reputation · Human in the loop
Answering Google reviews matters to every owner and nobody has an evening left for it. This drafts each reply and sends it to the owner in WhatsApp to approve.
VIII
Multi-agent systems
A query goes to a society of agents, not a search box: a broker routes it, an ontology agent widens it, two resource agents answer from Wikipedia and the web.
01 / 08
Scroll
