We've raised a €2M pre-seed · EU-Startups →

AI platform for Food & Beverage

Your R&D remembers
every recipe

Mafer turns your quality, formulation and regulatory records — specs, CoAs, trials, sensory panels — into model-ready intelligence, embedded in your daily workflows.

The platform
Recipes Lifecycle Agents Catalog Saved · v1

Recipe name

Dehydrated vegetable stock — reduced sodium

CAL-VER-2026-0xx · draft · v01

CategoryLiquid stock Target cost€1.90/kg Current€1.36 · −29%

Raw materials

All production sites
AllSaltsUmamiVegetablesFats & oilsCarriers
Monosodium glutamate (E621) €2.10/kg non-Ecolabel
Sodium ribonucleotides (E635) €14.80/kg non-Ecolabel

Composition

↕ Auto-order · Total 100.0%
Raw materialFunction%€/kg
Micronised vacuum salt Sodium contribution and base savoury profile 0.05
Food-grade potassium chloride Partial sodium substitute in reformulation 0.16
Maltodextrin DE 18 Flavouring carrier and body adjustment 0.13
RSPO SG palm vegetable fat Fat carrier and binder of the cube 0.12
Powdered yeast extract Clean-label umami; replaces declared glutamate 0.38
Dehydrated onion, 5 mm dice Aromatic base for stocks and sofritos 0.17
Dehydrated diced carrot Vegetable sweetness and colour in soups 0.15
Natural hydrolysed chicken flavouring Sensory signature of the chicken stock 0.19
Demineralised process water filler Carrier and balance adjustment 25.5 0.00
Total100.0%€1.36

Tip: the filler auto-completes to 100% — every panel recomputes live on each dosage change.

Mafer AI — draft v02 ready: sodium −28% vs current recipe, EU 1169/2011 labeling clear, taste profile preserved.

Taste profile

Savoury 100%
Fatty 20%
Sweet 18%
Vegetable 17%
Umami 15%
Bitter 11%

Each accord is drawn relative to the strongest — recomputed on every dosage change.

Cost

€1.36 per kg

per portion · dosage 8 g

target €1.90 · SAP standard cost · EUR @ 1

Nutrition · per 100 g

Energy124 kcal
Fat9.1 g
Sodium−28% vs current
Salt equivalent31.4 g

EU 1169/2011 · label auto-generated

Agents

formulation-v2 scored 96 candidates1.1s
labeling checked 14 markets · clear3.6s
quality within spec range · 24 batches0.8s
model mafer-food-v2 · fine-tuned on 8,912 of your recipes · last retrain 02:47 3 agents active · 240 ms
Book a demo 30 minutes · on your own use case

Years of specifications, trials, batch controls and market approvals become the foundation your teams — and your models — run on.

Quality & QC

Specifications, certificates of analysis, batch controls and non-conformities structured into one auditable quality system.

NPD & Reformulation

Historical trials, recipes and sensory feedback become reusable knowledge that shortens development and reformulation cycles.

Regulatory & Labeling

Ingredient rules, allergens, claims and market-specific requirements encoded into workflows that keep every label compliant.

Food companies run quality across spreadsheets, PDF certificates, ERP entries and lab notebooks. Answering "why did this batch deviate?" means hours of manual archaeology.

Mafer unifies specs, supplier documents, analytical results and batch records into structured, traceable data — so audits, deviations and supplier reviews take minutes, not days.

Cost pressure, ingredient substitutions, clean label, nutritional targets: every brief today restarts a cycle of trial and error your company has already paid for.

Models trained on your own history propose candidates that respect taste, process and compliance constraints — so your technologists iterate from a head start, not from scratch.

From ingredient and flavour houses to finished-product brands, these businesses share the same foundation: complex recipes, demanding quality standards and highly specialized data.

Book a demo

Quality Data Capture

Specs, CoAs & batch records

Full Traceability

From supplier to shelf

Predictive NPD

Models trained on your history

Market Compliance

Labeling & regulatory, always current

Very little of it becomes operational intelligence. The knowledge is there — it just doesn't compound:

Scattered Quality Data

Specs, CoAs and batch records live across spreadsheets, PDFs and inboxes.

Slow NPD Cycles

Every brief starts from scratch because past trials are not reusable.

Regulatory Drift

Labeling and market rules change faster than manual processes can follow.

Sensory Gap

Panel and consumer feedback never connects back to formulation data.

Eight weeks, on your own data. We start from plain exports, so there is nothing to integrate with your ERP, and we agree on the KPIs before the pilot begins.

Weeks 1–2

Data audit

We map what you have — ERP records, Excel macros, instrument outputs, scanned notebooks — and tell you exactly what's usable. No cleanup required on your side.

Weeks 3–6

Build

We build the proof on your own data — structuring it, automating the workflow, or training a model on your history. Whatever your case calls for, your data stays exclusively yours.

Weeks 7–8

Validation

Your team tests it on your own recipes, against the KPIs agreed upfront — before any long-term commitment.

≈3 h/week from your team Data residency of your choice Private deployment available
Book a demo We'll scope it on your data

FAQ

Real questions from R&D, IT and quality teams across our conversations — answered straight.

No. Each client gets their own models, fine-tuned exclusively on their own data. Nothing is pooled across clients, and nothing you share with us ever improves anyone else's system. What remains after training is your model — raw data is never reused beyond it.

Wherever you decide. We deploy in the region each client determines, whether that is Southern Europe or any other location, and we work with all major cloud regions. Data lives on enterprise-grade infrastructure, encrypted in transit and at rest. For companies with stricter policies we deploy inside your own cloud environment, and on-premise for the strictest ones. Our security architecture documentation is available on request.

That's the starting point for almost every client. Every engagement begins with a data audit: we map what exists — ERP records, spreadsheets with macros, instrument outputs, even scanned notebooks — and tell you exactly what's usable before any commitment. You don't need to clean anything first; structuring messy technical data is precisely what Mafer is built for.

A typical pilot runs about eight weeks: two of data audit, four of building on your data, two of validation against KPIs we agree on upfront. It runs on plain exports, with no ERP integration to start, and asks around three hours per week from your team.

Less than you'd expect. We start from base models built for your domain and fine-tune them on your history, so even small datasets move the needle. Quality beats volume: complete iteration records — trials, failures, adjustments — teach a model more than thousands of final formulas alone.