Nalu AI
PREDICTIVE & GENERATIVE AI

Ride the wave
of demand.

Nalu [nah-loo] Hawaiian for wave.

Born from a deep connection to the Pacific, years of experience underwater, and the conviction that data follows the same patterns as the ocean: seasonality, trends, noise — waves.

Spotting the right patterns, forecasting the next wave and being ready for it — in the ocean as in the data.

Built by Maximilian Fischer

Physics & Meteorology (LMU Munich) · ML Engineering · Munich

Building data-driven systems since 2016.

PROJEKT · 01

MeatMind

End-to-End ML & BI Platform

Independently developed · Full-time role · Müller Fleisch GmbH

Fully bespoke platform for demand forecasting, business intelligence and operational planning in meat processing. In production at a mid-sized manufacturer with 770+ customers and 900 SKUs.

The platform processes more than 450,000 time series data points and combines ML-driven demand forecasts with classical BI — from SAP R/3 data ingestion to ready-to-act decision support for sales and procurement. Generative AI (Ollama, on-premise) powers automated report generation and a natural-language data chat. The architecture is designed for both on-premise and cloud environments.

Modules

Demand Forecasting

ML-driven demand forecasts at SKU and customer level. Models: Temporal Fusion Transformer (PyTorch), LightGBM. Cold-start handling for new SKUs via segmentation-based global models.

Production Planning

Automated planning based on forecasts. Integrates capacities, minimum lot sizes and production cycles.

Slaughter Analytics

Analysis of slaughter data to optimise purchasing and production. Links carcass weights, quality data and demand forecasts.

Procurement Forecasting

Forecast-driven purchasing for live cattle — which species, husbandry type and quantity to source, derived from demand forecasts and production capacity.

Customer Segmentation

ABC/XYZ analysis at customer and SKU level. Dynamic segmentation as the foundation for forecasting and sales steering.

SCM Module

Supply chain overview with inventory trends, supplier performance and bottleneck detection.

AI Chat

Natural-language access to platform data and analytics. LLM-driven queries on demand, customer and SKU data. Powered by Ollama (on-premise, no data leaves the server).

Automated Report Generation

LLM-generated reports based on current forecasts, segmentations and operational KPIs. Context-aware summaries instead of static templates.

Further Developments

ETL & Data Automation

Fully automated pipelines (Dagster/Airflow) for SAP R/3 data extraction, transformation and loading into the analytical data warehouse.

Internal Tools

Cattle ear-tag lookup with API integration to Qualifood — real-time queries on origin and quality data directly from the platform.

Technical Challenges

SAP R/3 Data Quality

Automated detection and correction of mis-postings, inconsistent master data and duplicate entries.

Scale

Training and serving 450,000+ time series efficiently — solved via segmentation-based global models instead of one model per SKU.

Cold-Start Problem

Forecasting new SKUs with no historical data — handled via assignment to existing segments and transfer learning.

REFERENCE

"The applications built here are now a fixed part of our daily operations."
Managing Director · Food manufacturer · German mid-market

Numbers

0

customers

0

SKUs

0

time series

0

weeks to go-live

Stack

PyTorchLightGBMFastAPIReactTypeScriptDuckDBDockerNginxMLflowDagsterAirflowCeleryRedisSAP R/3Microsoft Entra ID (SSO)RBACCI/CDSHAPOllama
PROJEKT · 02

Nalu AI

Demand Intelligence Platform — Next Generation

Independently developed

Generalisation of MeatMind into an industry-agnostic platform. Multi-tenant architecture with configuration-driven onboarding — no hard-coded industry assumptions in the core.

Each customer receives a dedicated instance, deployed via Docker Compose on their own infrastructure. No customer data leaves the server.

app.nalu-ai.com/dashboard

Übersicht

KW 18 · Mo, 8. Mai

7T30TYTD

Gesamtabsatz

€ 2.4 M

+8.2 %

Forecast-Genauigkeit

96.2 %

MAPE 7.6 %

Service Level

98.1 %

+1.4 %

Aktive Forecasts

894

12 neue

Absatz vs. Forecast · 12 Wochen

LightGBM · MAPE 7.6 %

Historie Forecast
HEUTE

ABC/XYZ-Matrix · 894 Artikel

automatisch klassifiziert

A
B
C
X
142
96
48
Y
78
134
112
Z
22
64
198
A · Hoher Umsatzanteil
B · Mittelwert
C · Geringer Umsatz
XYZ = Variabilität · niedrig → hoch

Dashboard with KPIs, forecast chart and ABC/XYZ matrix.

Architecture Highlights

Multi-Tenancy

Every customer is a dedicated instance with its own configuration. Onboarding through config.yaml — no code changes required.

Zero-Industry-Assumptions

Core code carries no industry logic. Column names, modules and features are fully configuration-driven.

On-Premise or Cloud

Docker Compose on own infrastructure or cloud deployment. Flexible per requirement — no vendor lock-in, full control over data and operations.

app.nalu-ai.com/forecasting

ARTIKEL

Nordpils Premium 20×0,5L

BIE-001 · 8-Wochen Forecast

Modell · LightGBM
MAPE · 8,4 %
Accuracy · 91,6 %

Absatz-Forecast · KW 38 – KW 46

Historie Forecast
HEUTEKW 38KW 39KW 40KW 41KW 42KW 43KW 44KW 45KW 46

312

P10 · pessimistisch

387

P50 · erwartet

463

P90 · optimistisch

⚡ Empfohlene Bestellung: KW 41 · Menge: ~1.550 Einheiten

Letztes Training: heute 04:12 UhrNächstes Retraining: Mo 04:00 Uhr

SKU forecast with P10/P50/P90 intervals and model metrics.

app.nalu-ai.com/scm/alerts

Reorder Alerts

heute · 14:32

2 kritisch2 Warnung
ArtikelBestandROPReicht

Salami Mailand 80g

4711-Salami

54

238

2 T

kritisch

Cola Classic 1L

6812-Cola

112

420

3 T

kritisch

IPA Craft Series 0.33L

9023-IPA

287

360

6 T

Warnung

Bio Müsli 500g

3401-Müsli

198

240

8 T

Warnung

Tomaten passiert 400g

7765-Tomaten

642

580

14 T

auf Soll
5 von 894 Artikel · gefiltert: nicht-OK↻ Push: Slack · Teams · E-Mail

Reorder alerts with severity classification and push notifications.

app.nalu-ai.com/kunden

KUNDEN-INTELLIGENZ

Wer kauft wann — und was.

Kaufbereit (30T)

143

Erw. Umsatz

€184k

Churn-Risiko

12

KundeKaufwahrsch.TerminStatus

REWE Group

KD-001 · Champion

92%

KW 20

Kaufbereit

Edeka Südbayern

KD-002 · Loyal

78%

KW 21

Kaufbereit

Gastro Service Nord

KD-047 · At Risk

34%

Churn-Risiko

Metzgerei Huber

KD-089 · Loyal

61%

KW 22

Bald kont.
770 Kunden · aktualisiert täglich 06:00 · Modell: Survival Analysis

Customer intelligence with purchase probability and churn risk.

Stack

PythonFastAPIPyTorchLightGBMReactTypeScriptDuckDBPostgreSQLDockerCeleryRedisNginxMLflowAlembicstructlog
PROJEKT · 03

E.ON

Data Quality & Risk Engineering

Independently developed · Full-time role

End-to-end quality assurance of calculated price data — fully automated in Python. Validation of regulatory price adjustments and procurement costs in energy trading.

Analytical framework for risk assessment (Initial Margin) in commodity portfolio management.

Stack

PythonSQL
PROJEKT · 04

Dive Operations Platform

Full-Stack Web App

Independently developed · 2 years · Hawaii · PADI Dive Master

Full digitisation of a dive shop on Hawaii during a sabbatical. All operational processes — bookings, customer management, daily planning, certification tracking — migrated from paper and whiteboard to a single web application.

Before

Paper sheets, whiteboards, Excel, manual coordination

After

One app for bookings, CRM, daily planning and certifications

STACK

ML & Data Science

PyTorchscikit-learnLightGBMSHAPMLflow

Generative AI

OllamaLLM-driven report generationAI chat

Backend & APIs

PythonFastAPIFlaskCeleryRedis

Orchestration

DagsterAirflow

Data & Storage

DuckDBPostgreSQLSnowflakeSQL ServerParquet

Frontend

ReactTypeScript

Infrastructure

DockerKubernetesNginxGitHub ActionsCI/CDPrometheusGrafana

Cloud

AzureEntra ID (SSO)Azure Blob Storage

Integration

SAP R/3REST APIsstructlogCustom Observability