RECIPE + DOSING EXECUTION
Order of addition and dosing control per vessel, with tolerance-aware compensation when a raw-material lot behaves differently from the last one.
STAGE 01 // MAKE
Emulson
makes the batch.
The make brain of Serumon. Emulson executes recipe and order of addition, controls raw-material dosing, and actively regulates temperature, shear, vacuum, homogenisation pressure and time to build target viscosity, stability, particle size and feel — using a learned process world model of each formula and vessel.
1–10 HzSTATE ESTIMATION
MPC + RLCONTROL POLICY
PER FORMULAWORLD MODEL
DGX + ORINGPU PROFILE
$ emulson trace --last 6 t+12:04 ADD PHASE B OK dose 148.2 kg (-0.3%) t+13:20 TEMP HOLD 78.0C OK drift 0.14C t+15:02 HOMOGENISE 3200 rpm TRIM -40 rpm x 90s t+16:41 D50 PREDICT 2.24 um IN SPEC t+18:42 VISC PREDICT 41.8k IN SPEC t+19:10 STABILITY RISK 0.06 (low) reason: shear trim taken to protect D50 while viscosity tracked the upper envelope.
RECIPE + DOSING EXECUTIONEMULSION STATE ESTIMATIONMODEL-PREDICTIVE CONTROLFORMULA WORLD MODELBOUNDED ENVELOPESGMP AUDIT LOG
THE PROBLEM
A drift you cannot see breaks a batch you cannot rescue
A cosmetics batch is built in a jacketed vessel by dosing oils, waters, actives, pigments and surfactants in the right order, then emulsifying and homogenising under exact temperature, shear, vacuum and time.
A small drift in any of those breaks the emulsion, shifts viscosity or pH, coarsens particle size or ruins sensory feel. By the time the lab result lands, the batch is already made — so it gets held, re-worked or scrapped.
The compounding chemists whose senses govern this black art are scarce and ageing, and most plants still run manual recipes with reactive lab QC.
TYPICAL LOSS PER HELD BATCH
- RAW MATERIAL
- $18k–$90k
- VESSEL TIME
- 6–14 h
- RE-WORK LABOUR
- 2–5 shifts
- LATE DELIVERY RISK
- HIGH
- ROOT CAUSE FOUND
- RARELY
ILLUSTRATIVE SITE FIGURES [ASPIRATIONAL UNTIL MEASURED].
CAPABILITIES
What Emulson does
EMULSION STATE ESTIMATION
Real-time viscosity, particle size, pH proxy and stability-risk estimation from inline sensors — hours before the lab confirms it.
MODEL-PREDICTIVE CONTROL
MPC over temperature, shear, vacuum, homogenisation and time that steers the batch to spec instead of holding a fixed recipe.
FORMULA WORLD MODEL
A per-formula, per-vessel model with plant memory: it knows how this vessel behaves on this product at this fill level.
BOUNDED ENVELOPES
Every actuatable parameter is clamped by envelopes signed off by process and quality, with human-in-the-loop checkpoints.
GMP AUDIT LOG
Immutable record of every prediction, trim and approval, replayable against the exact model version that ran the batch.
THE LOOP
Four beats, every batch
-
01
LOAD FORMULA
Recipe, spec envelope and vessel profile are resolved, with the raw-material lots for this batch and their historical behaviour.
-
02
DOSE + ADD
Dosing is executed and verified per phase, compensating within tolerance for lot variability and load-cell drift.
-
03
EMULSIFY
Temperature, shear, vacuum and time are actively regulated while the world model predicts where viscosity and D50 will land.
-
04
HAND OFF
The batch is released to fill only when predicted spec, Shadera shade and Stabion risk all agree. Everything is signed into the audit log.
AI CAPABILITIES
What is real today, and what is not
Serumon labels its own maturity. Anything not yet validated in production is marked aspirational, on this site and in the room.
REAL
PROCESS WORLD MODEL
Multimodal model over inline sensors, formula structure and batch tags. Real for design-partner formulas; fleet-scale generalisation is aspirational.
REAL
RHEOLOGY PREDICTION
Viscosity, particle-size and emulsion-fault prediction validated against lab results on the wedge formula family.
PARTLY ASPIRATIONAL
RL / MPC CONTROL
Shear-time-temperature-vacuum control policy. Runs in recommend mode from day one; write access per site follows validation.
REAL
GROUNDED REASONING
Frontier-model reasoning over SOPs and batch records with enforced citations for deviation explanation and formulation Q&A.
NVIDIA ALIGNMENT
GPU is must-have, not nice-to-have
DGX / HGX
Train process world models and control policies over spectra, telemetry, formulas and batch records.
JETSON ORIN / IGX
Plant-edge inference and the control loop at each compounding vessel, with deterministic latency.
CUDA / TENSORRT
Quantised inline inference to hold p95 50–150 ms control decisions.
MODULUS + NEMO
Physics-informed emulsification surrogates, plus a fine-tuned cosmetics-formulation reasoning model.
MEASUREMENT
What we get judged on
RIGHT-FIRST-TIMEPRIMARY METRIC
SCRAPSECONDARY METRIC
VESSEL HOURSTHIRD METRIC
$9k/moLINE TIER PRICE
PILOTS ARE CONTRACTED AGAINST A NAMED ROI METRIC AND A BASELINE WE RECORD BEFORE ANYTHING CHANGES.
THE MOAT
Legacy control automates sequences. Emulson controls chemistry.
MES, DCS and mixer OEMs automate the sequence: hold this temperature, run this rpm for this long. None of them model the physical chemistry the sequence was supposed to produce.
Emulson closes the loop on the actual emulsion with a learned per-formula world model, delivering right-first-time batches no PID recipe can match — and every chemist correction makes it harder to displace.
- Proprietary per-formula and per-vessel process traces.
- Chemist corrections captured as labelled training data.
- Plant-specific memory of vessel and lot behaviour.
- Switching cost = losing the captured craft.
DEPLOYMENT SHAPE
- TIER
- LINE / PLANT / ENT
- LINE PRICE
- $9,000 / mo
- PLANT PRICE
- $60,000 / mo
- ENTERPRISE
- $400k–$5M ACV
- EDGE NODE
- PER VESSEL CLUSTER
- AUTONOMY
- GRADUATED
ECOSYSTEM
How EMULSON plugs into the loop
Every agent hands the next one a better decision. Buy one; the others make it sharper.
SHADERA
undefined
FILLIX
undefined
STABION
undefined
TWYNEX
undefined
EMULSON FAQ
Straight answers
One formula family on one vessel. We pick the family with the highest scrap or re-work rate and enough batch frequency to produce a statistically useful baseline inside the pilot window.
Often some. Most vessels already have temperature, pressure, vacuum, load cells and drive telemetry. Inline rheology or particle sizing may need a retrofit; sensor retrofit cost per vessel class is scoped during the site survey.
Emulson writes setpoints, not safety logic. Your DCS and interlocks remain authoritative, and every write lands inside an envelope your process owners approved.
Prediction accuracy versus lab results, right-first-time rate, scrap and re-work volume, and vessel hours recovered — all measured against the pre-pilot baseline we recorded in shadow mode.
WEDGE PILOT
Run Emulson on one line
Instrument, baseline, shadow, then assist — against a metric you name before we start. If the numbers do not move, you have lost a quarter of telemetry and gained a very detailed picture of your own process.