Maximilian Fischer, AI Engineer
Ride the wave of demand.
Nalu is Hawaiian for wave.
Since July 2025 I have been building and running, as the sole developer, the internal forecasting and planning platform of a mid-sized meat industry group with three sites. The data platform behind it, the replacement of the integration platform and the leadership of the AI team being set up are mine as well.
What I claim, I back with measurement. Including where the result is uncomfortable.
Key figures
Each number has a project behind it, with measurement and source.
Count of documented replacements, as of October 2026: 13 up to July 2026, since then the daily list of open items and seven Excel reports from livestock settlement. An entry counts only once the new path is in use.
Weighted mean absolute percentage error (WMAPE) at the animals-per-week level, walk-forward over 159 weeks out of sample, against the previous-week carry-forward at 28.2 %.
Line count of the platform’s source code, rounded, as of July 2026.
Sum of rows from the four source systems in the data platform, ten years of history, rounded.
Five projects, each with the same structure: starting point, decision, implementation, result with measurement, and what I would do differently.
An internal platform for forecasting and planning, built and run by one person. Now in use at all three sites: 60 people in 90 days, 54 active people in September 2026.
Walk-forward over 159 weeks against the previous-week carry-forward. At the decision level, 20.6 instead of 28.2 percent error. At the top aggregate level the simple carry-forward is marginally better, and that is stated here, not in a footnote.
57 assets, 16 jobs, around 164 orchestrated steps per working day over around 155 million rows from four source systems. Hardened after real incidents. Third site connected, with peak memory brought down from 39.3 to 9.7 gigabytes.
Ongoing. Inventory captured automatically: 99 scheduled processes, 1,278 jobs, around 8,200 processing steps. As of October 2026 all 99 processes have been worked through, 96 of them replaced, built or deliberately dropped. 19 already run in parallel to the original.
Ongoing, concept phase. A digital visitor process for all sites, from request to deletion. A rule set of 11 visitor types and 9 zones determines briefing, protective clothing and approval. The connection to the existing access control system is verified from its database, not assumed.
A daily briefing for every user, a short profile per customer and a weekly report per salesperson, written by a local language model. Every number in the text is checked against the data. And when AI agents reviewed the platform, I checked the reviewers: 3,258 statements, 82 of them wrong.
How I work
Method instead of a tool list. Three of eight points:
I do not start with the target picture but with what exists.
Before I build anything, I fix what it competes against.
The level on which the model loses goes in the table, not in the footnote.
Own venture, in the making
Decision platform for demand forecasting, configurable per customer, every instance at the customer. The entry point is a list of decisions that are due, not a wall of metrics.
In the making, not an offering. Drafts with sample data, no numbers until they are measured.
Today
Week 37, Monday, 7 September 2026, model run 05:12
Forecast, next 4 weeks
Nalu expects 246,000 crates, +9.4% against the same period last year, range ±21%. The increase comes almost entirely from the heatwave and the end of the holidays. Three decisions are due, one of them before Wednesday.
Mineral water Classic lasts 6 days. The spring delivers in 8.
The four water items pull 14,200 crates next week, 18,900 at peak. Stock covers until 13 Sep, empties returns cover 70% of the volume. The order would have to go out today.
Week 38 exceeds picking capacity with 68% probability
Expected 61,400 crates against a capacity of 54,000. Even in a normal week, 7,400 crates are missing. Pulling forward into week 37 solves it. 9,200 crates are free there, the warehouse can take it.
Hotel Seeblick has ordered below forecast for three weeks
Expected 320 crates a week, actual 190. Well below what even a bad week explains. So not a coincidence. Churn 58% likely within 60 days, annual value €36,000.
last 30 days, off by 8.1% on average, slightly too low
From 8 weeks on, plan with the range, not with the single number.
118 of 124 items run inside the forecast corridor. Availability 100%, no open alerts.
Technology
Tools from the projects, without rating bars. How I used each of them is described with the respective project.