VD
Logbook21 in 12 months
DeMaVision47 cameras, one sensor
Friday’s day 
Sibyl160,149 sales lines
Elevate Fresh50 KB, down from 256

Vincent Demarez

I build the software our family business runs on, and the AI that runs my day.

Systems I built and run

Most of these started with a problem on the factory floor at Agrafresh, the fresh-cut produce company my parents founded. The rest started with a problem in my own week.

2025–2026

Friday

An AI chief of staff that lives in iMessage

Friday reads my messages, mail and calendar, and acts on them. It plans my day, chases people who still owe me an answer, briefs me at seven, and picks up when I call it from the car. It started as an HR-law assistant I co-built at school and grew into the system I lean on most.

Built for
Me, every day
Runs on
A home Mac hub, a laptop worker and my iPhone
Stack
Python, Swift, CallKit, vector memory, LLM routing
Status
Live
How Friday works
FFriday

Today 07:00

Morning. Light Tuesday: two calls, and 13:00–15:00 is blocked for deep work.

Leave by 07:40. Traffic is building on the E40.

Two threads are waiting on other people: the supplier quote (day 3) and Saturday’s venue.

Nudge the supplier

Drafted: “Hi Tom, any news on the quote? Happy to jump on a call this week.” Send it?

ja

Sent. I’ll tell you when he answers.

FFriday02:14

YouWhat’s still open today?

FridayTwo things. Sign off the quality report, and call the packaging supplier back.

Recreated with invented content. Friday’s real threads are private.

2026

Sibyl

Demand forecasting for Agrafresh

Sibyl predicts how much of each product our retail customers will order, so the factory can plan cutting and staffing before the orders land. It learned from 160,149 sales lines over three and a half years, and it retrains itself every night.

Built for
Production planning at Agrafresh
Runs on
Our own server
Stack
Python, TimesFM, LightGBM, Optuna
Result
34.1 WAPE for TimesFM against 36.5 for LightGBM (lower is better)
How Sibyl works
0.250.500.751.00same week last year41 product lines, best first0.250.500.751.00same week last year41 product lines, best first

Typical product line: 32% less error than the baseline. Where nothing beat it, Sibyl keeps the baseline.

TimesFM
15
Ensemble
11
LightGBM
7
TimesFM with signals
4
Last year’s week
4

2026

DeMaVision

Production intelligence from camera footage

The factory already had 47 security cameras. DeMaVision turns them into a production sensor: it sees when a line runs, where it stalls and when packing waits for product, then checks that against the ERP. No new hardware, and no footage ever leaves the building.

Built for
Production at Agrafresh
Extra hardware
€0
Stack
Python, computer vision, Hikvision recorders
Status
Measuring
How DeMaVision works
  1. Cameras47 existing cameras on three recorders
  2. Vision engineRuns locally, on our own server
  3. ERP cross-checkWhat the cameras see, matched with what the system says
  4. Line stateRunning, stalled, or starved for product
  5. Daily reportWhere the time went, and why

2025–2026

Libra

Workforce planning for the factory floor

Libra brings the competence matrix, the shift roster, evaluations and candidate pools into one view, so HR and production can see who can run which station, on which shift, and where the real gaps are. It never pretends someone is available when they aren’t.

Built for
HR and production at Agrafresh
Generation
Third: two Streamlit prototypes, then a full rebuild
Stack
React, TypeScript, Convex
Status
Live
How Libra works
  1. Competence matrixWho can do what, and how well
  2. RosterWho works which shift
  3. EvaluationsHistory, not hearsay
  4. Candidate poolsInternal and external
  5. Staffing per shiftThe real gaps, with no invented availability

2026

Aletheia

Quality control, off paper

Aletheia moved two paper quality processes onto a tablet: inspecting incoming raw material, with a supplier report and a 24-hour window to dispute it, and the daily checklists on finished product. Reports build themselves as PDFs and go out by email.

Built for
The quality team at Agrafresh
Stack
React, TypeScript, Convex, react-pdf, Resend
Runs on
Our own server, moved from Vercel
Status
Live
How Aletheia works
Aletheia
Real screen. Names and figures blurred.

2026

Mesh

A group chat for AI agents

Mesh is a group chat for personal AI agents. Your own machine hosts the room; you send a link, and a friend’s agent, whether that’s Grok, Hermes, ChatGPT or Cursor, walks in and talks. No app, no account, and nobody in the middle.

Built for
People who run their own agents
Stack
Hono, MCP, Bun
Runs on
Whoever hosts the room
Status
Live at trymesh.chat
How Mesh workstrymesh.chat
trymesh.chat

Websites

Two sites, each built for one audience: people deciding what to eat, and people deciding where to get their medicine.

Elevate Fresh

Elevate Fresh

Brand site for our line of chilled dressings

Drafted in a design tool, then rewritten by hand as a 50 KB page so it paints at once on a phone. The signature pour runs on the compositor only, so it stays smooth while you scroll.

Visit elevate-fresh.com
Apotheek Kerckaert

Apotheek Kerckaert

Site for a neighbourhood pharmacy in Zedelgem

Opening hours, services and vaccinations in plain Dutch. Every page also exists as clean markdown with an llms.txt, so AI assistants answer questions about the pharmacy correctly.

Visit apotheekkerckaert.be

Logbook

Everything else, newest first. Prototypes are marked as prototypes.

DateProjectBuilt withStatus
Internship matcherSwipe through internships scored against your CV; it learns your taste from the swipes.React, Convex, Cloudflare WorkersOpen source
Transport app for a logistics clientFreelance: a Flutter app for a transport company’s drivers and planners.FlutterIn progress
Deal HunterFinds second-hand deals in East and West Flanders and mails the ones worth my time. Free scans four times a day; paid AI scoring once, on new items only.Python, SQLiteRuns daily
orchestrate-skillA Claude Code skill that plans a goal and hands each piece to the best coding agent installed.Shell, MarkdownOpen source
Field consoleHarvest, irrigation and scouting for our fields, with growing-degree-day targets per crop and a daily AI agronomist brief.JavaScript, Supabase, Open-MeteoPrototype
SenecaA life operating system: tasks per company, an auto-planned calendar and my agents’ status on one screen.React, ConvexPersonal
Appartement HunterScrapes the Belgian property sites, scores each listing’s condition and feel with a vision model, and mails a ranked digest every morning.Python, vision LLMShipped
EasyncMy small web and automation studio. It started as my bachelor thesis, which won best thesis.Web, Apps Script, PythonRunning
Vacancy monitor for a recruitment agencyWatches career pages, spots new and removed vacancies, and sends the agency a daily digest in Dutch.Python, PlaywrightClient work
AereliusA working simulation of payments for AI agents: budgets, velocity limits, merchant allowlists, human approval, virtual cards and an audit trail.JavaScriptPrototype
SommelierPhotograph a wine label; it identifies the bottle and advises from its own curated wine knowledge.Streamlit, Groq vision, SupabasePrototype
iMessage exporterA desktop app that finds your Messages database and exports threads to JSON, on macOS, Windows and Linux.Tauri, Rust, Next.jsReleased
PulseCheckWeekly engagement surveys for teams: AI writes the questions and reads the answers, and results stay anonymous.Next.js, Prisma, PostgresPrototype

About

I’m Vincent, 23, from Ghent. I grew up around Agrafresh, the fresh-cut produce company my parents founded in Pittem, and most of what I build starts there: a job on the floor that software can take off someone’s hands.

By day I’m a technology and transformation consultant at EY, working on SAP SuccessFactors. Before that I did a Master in Innovation & Entrepreneurship at Vlerick and an MBA at VUB.

I build with AI agents as my team and stay strict about what ships: real numbers, software that tells the truth, and nothing that only works in the demo.

Specifications

Based in
Ghent, Belgium
Works at
EY, Technology & Transformation
Family business
Agrafresh, fresh-cut produce
Studied
Vlerick (MIE), VUB (MBA)
Speaks
Dutch, French, English

Got a process that should run itself?

I like problems with a real operation behind them. Tell me what yours looks like.

Email me

demarezvincent@gmail.com