MAKE
EMULSON
Executes recipe and order-of-addition, controls dosing, and regulates temperature, shear, vacuum, homogenization and time to build target viscosity, stability, particle size and feel.
PHYSICAL AI // COSMETICS MANUFACTURING
Every batch,
perfectly on shade.
Serumon runs the cosmetics plant: autonomous compounding and emulsification, inline shade and defect vision, fill-weight control with no giveaway, predictive micro and stability assurance, and a make-and-fill digital twin. One closed perceive → decide → act loop at the plant edge.
5AUTONOMY AGENTS
$20BTAM
ISO 22716GMP-ALIGNED
<150msEDGE CONTROL P95
$ serumon status --vessel V-07 FORMULA SPF50 EMULSION / REV 14 STAGE HOMOGENISE t+18:42 VISCOSITY 41,800 cP target 40-44k PARTICLE D50 2.1 um target <2.6 SHADE dE 0.42 hold <0.80 STABILITY LOW RISK conf 0.94 > emulson: shear -40 rpm for 90s (approved envelope) > shadera: no drift detected, fill gate OPEN > stabion: release evidence pack 82% assembled
AUTONOMOUS COMPOUNDINGSHADE ΔEZERO GIVEAWAYISO 22716EMULSION STABILITYROBOTIC PACKOUTBATCH RELEASEDIGITAL TWINPLANT-EDGE GPUFIRST-PASS YIELD
THE PROBLEM
A cosmetics batch is a black art running on eyeballs
Serums, creams, foundations, lipsticks, shampoos and sunscreens are precisely engineered emulsions, suspensions and dispersions that must look, feel, spread and perform identically in every jar — and match a brand shade to an imperceptible tolerance, batch after batch, at scale.
A small drift in dosing, order of addition, emulsification temperature, shear or homogenisation pressure breaks the emulsion, shifts viscosity or pH, coarsens particle size or throws the shade. Because colour is judged by eye against a standard, whole batches get held, re-shaded or scrapped over a barely perceptible mismatch.
The compounding chemists, shade matchers and line technicians whose senses govern quality are scarce and ageing. Most plants still run manual recipes, eyeball shade matching and reactive lab QC.
WHERE THE MONEY LEAKS
- Emulsion and rheology drift breaking stability, viscosity, pH or particle size.
- Shade mismatch causing batch holds, re-shading and scrap.
- Overfill giveaway handed away to guarantee minimum fill weight.
- SKU proliferation and launches exploding changeover and cleaning validation.
- Micro contamination and stability failures triggering holds and recalls.
- Volatile, expensive pigment, active and packaging inputs plus re-work loss.
TARGET SITES CARRY $250K–$2M/YEAR OF ADDRESSABLE WASTE PER PLANT [ASPIRATIONAL UNTIL MEASURED IN A PILOT].
THE SERUMON LOOP
Five agents. One closed loop.
Make → perceive → fill → assure → simulate. Each agent feeds the next, and every batch trains all of them.
STAGE 01MAKE
Emulson doses, emulsifies and holds the rheology envelope in the vessel.
STAGE 02PERCEIVE
Shadera reads shade ΔE and cosmetic defects inline, at line speed.
STAGE 03FILL
Fillix trims filler setpoints, kills giveaway and runs the robot cells.
STAGE 04ASSURE
Stabion predicts micro and stability risk and assembles the release pack.
STAGE 05SIMULATE
Twynex twins the whole make-and-fill line and re-optimises the campaign.
THE PRODUCTS
The plant, product by product
PERCEIVE
SHADERA
Inline spectrophotometric ΔE prediction against the brand standard plus high-speed defect vision across compacts, bullets, jars, tubes and bottles.
FILL & PACK
FILLIX
Controls fill weight and level to strip out giveaway while holding compliance, and drives robotic assembly, capping, decoration and case-packing.
ASSURE
STABION
Fuses bioburden, preservative-efficacy and stability signals with make-and-fill process context into one predictive batch-safety and release picture.
SIMULATE
TWYNEX
Simulates compounding, emulsification, shade and filling to hit target spec before the batch, then optimises campaign sequencing, changeover and allocation.
RUNTIME
PLANT EDGE
One GPU runtime per vessel cluster and line, with deterministic latency, offline survival and OTA model delivery.
HOW IT RUNS
Perceive. Decide. Act. Prove.
The same four beats run on every vessel and every filling line, at 1–10 Hz, inside the plant.
-
01
PERCEIVE
Inline sensors, spectrophotometry, checkweighers and vision fuse into a live state estimate of the emulsion, the shade and the fill.
-
02
DECIDE
A learned per-formula, per-vessel world model predicts where the batch lands and computes the trim that hits spec.
-
03
ACT
Setpoints are written back into the vessel, homogeniser, filler and robot cell inside approved envelopes.
-
04
PROVE
Every perception, decision and action is signed into an immutable, ISO 22716-aligned audit trail with its evidence.
THE PRIZE
What autonomy is worth per plant
$20BTAM
Cosmetics manufacturing automation, batch/MES, inspection, line control and plant software.
~13%MARKET CAGR
Fastest growth in autonomous compounding, shade AI and fill optimisation.
$5BSAM
Plant-edge autonomy sold to brand owners, CDMOs and equipment channels.
$300M3-YR SOM
Narrow land-and-expand across line, plant and enterprise deployments.
MARKET FIGURES ARE SERUMON ESTIMATES; SOM IS [ASPIRATIONAL].
WHY NOT THE INCUMBENTS
They measure. Serumon acts.
MIXER OEMS
AUTOMATE SEQUENCES
Vessel and homogeniser controls hold a recipe. They are blind to the emulsion quality the recipe was meant to produce.
SPECTRO VENDORS
MEASURE COLOUR
A spectrophotometer tells you the batch is off shade. It does not tell the vessel which pigment dispersion to trim.
FILLING LINES
HOLD A SETPOINT
Checkweighers reject outliers at a fixed threshold. Giveaway keeps being handed away in the gap.
INSPECTION
CATCH DEFECTS LATE
Vision systems find the defect after the product is made, with no path back into the compounding decision.
MES / BATCH
RECORD HISTORY
Batch systems document what happened. They do not predict where the batch is heading or move a setpoint.
SERUMON
CLOSES THE LOOP
One plant-edge platform that perceives, decides, writes back into vessels, fillers and robots, and proves it under GMP.
HEAD TO HEAD
Point tools vs the system of action
| DIMENSION | INCUMBENTS / COPILOTS | SERUMON |
|---|---|---|
| SCOPE | A single task or tool | The full compound → emulsify → shade → fill → pack loop |
| LOOP | Setpoints and measurement | Closed perceive-decide-act at the plant edge |
| DATA MOAT | Limited, per-instrument | Cross-plant formula, shade and fill network |
| INTEGRATION | Read-only or shallow | Deep write-back into vessels, fillers and packaging |
| BUYER OUTCOME | Assist a human | Perform the work; humans supervise exceptions |
| LEARNING | Static thresholds | Every supervised correction retrains the world model |
GRADUATED AUTONOMY
Trust is earned one rung at a time
No plant hands over a vessel on day one. Serumon ships with an explicit autonomy ladder, per formula and per line.
RUNG 01
SHADOW
The agent observes the vessel or line and predicts. It writes nothing. You compare its calls against your chemists and your lab for the whole baseline period.
RUNG 02
RECOMMEND
The agent proposes the dose, shear, shade or fill move with its reasoning and evidence. An engineer accepts or rejects; every rejection trains the model.
RUNG 03
APPROVE-TO-ACT
The agent stages a control move inside an approved envelope. A named operator releases it. Every action is signed into the immutable audit log.
RUNG 04
BOUNDED AUTONOMY
Inside validated envelopes and on qualified formulas the agent acts on its own, escalating anything outside the envelope to a human within one control cycle.
WHAT BUYERS SAY
The room we sell into
Show me shadow mode on one vessel for a quarter, against my own chemists, and I will tell you whether it earns write access.
Our shade matchers are the best in the building and there are four of them for thirty lines. That is the actual constraint.
Giveaway is the number nobody puts on a slide. It is a rounding error per jar and a budget line per year.
All customer scenarios on this site are composite and illustrative. Serumon is pre-revenue; performance figures are design-partner targets, not audited results.
THE MOAT
The craft becomes the dataset
Every batch generates process traces, spectra, defect images, fill telemetry, micro and stability outcomes — and every chemist correction is a label. That pairing of physical process to expert judgement cannot be bought, scraped or licensed.
It compounds per formula, per vessel and per plant. A competitor starting later on the same site starts with none of it, and switching away means losing the captured craft of the people who already retired.
- Proprietary per-formula, per-vessel process traces.
- Shade spectra and ΔE outcomes tied to pigment lots.
- Defect image corpora across compacts, bullets, jars and tubes.
- Fill and robot telemetry against giveaway and compliance limits.
- Micro and stability outcomes linked back to make signatures.
DATA FLYWHEEL / PER SITE / PER QUARTER
- BATCH TRACES
- 1,200+
- SHADE SPECTRA
- 40,000+
- DEFECT IMAGES
- 2.1M
- FILL EVENTS
- 9.4M
- EXPERT CORRECTIONS
- 3,800
- MODEL REFRESHES
- 12
DESIGN-PARTNER SCALE TARGETS [ASPIRATIONAL].
FIRST QUESTIONS
What plants ask first
No. Every deployment starts in shadow mode with read-only telemetry. Write access is only requested after the agent has been measured against your own chemists and lab on your own formulas, and it is always bounded by envelopes your process owners define.
A wedge pilot runs on one vessel or one filling line. Typical shape: 4 weeks instrumentation and baseline, 8 weeks shadow, then assist mode against a named ROI metric such as shade-hold reduction, giveaway reduction or first-pass yield.
It does not have to. Serumon runs on plant-edge GPUs with on-prem inference, per-tenant model scoping and formula IP protection. Cloud training is optional and contractually scoped.
Every prediction ships with a confidence and its evidence. Below threshold, the agent escalates instead of acting. Actions are reversible inside the envelope, and the audit log records who approved what.
START NARROW
Put one vessel on Serumon
Pick your worst shade-hold SKU or your highest-giveaway filling line. We instrument it, run shadow mode against your own baseline, and show you the numbers before anything writes a setpoint.
DESIGN-PARTNER PROGRAMME — 3 TO 5 PLANTS.