Parith Thiengtham
Yale graduate. Founder of Edsy, an award-winning, profitable AI edtech startup. LinkedIn
Bangkok, Thailand
From petroleum to plants
A deep learning system that turns variable natural feedstocks into advanced materials for data centers, EVs, robotics, and semiconductors.
The problem
Petroleum wins on consistency, not merit.
The industries building the future run on the chemistry that is poisoning the planet — every AI data center, every EV, every semiconductor depends on petroleum-based resins, seals, and adhesives.
Feedstocks vary batch to batch, but precision manufacturing cannot. That mismatch has kept petroleum the default for sixty years — not because it's the better material, but because it's the more consistent one.
So we solve the variability, not the chemistry.
The feedstock varies. The material does not.
Ai-chemi
A control system for messy inputs.
Ai-chemi senses what a batch of natural feedstock actually is, predicts what it would become under a given process, and then works backwards — solving for the process settings that land the finished material on spec. Two models, pointed in opposite directions. The feedstock and the target specification are both inputs, not assumptions built into the architecture.
Forward model
What will this batch become?
Predicts finished-material properties from the sensed state of the feedstock and the process settings applied to it.
feedstock state + process settings
↓
predicted properties
Inverse controller
What should we do about it?
Takes the property targets a customer specifies, reads the batch in front of it, and returns the recipe that hits those targets.
target properties + sensed feedstock
↓
process recipe
Both models are physics-informed: cure kinetics and network-property relations are built into the architecture rather than learned from scratch. That constraint is what makes the system trainable on the number of runs a real pilot plant can produce, instead of the number a simulator can.
A working prototype already predicts adhesive properties — bond strength, cure kinetics, glass-transition temperature — directly from raw feedstock composition. We walk partners through it directly rather than publishing it. Get in touch to arrange a demonstration.
What we make
Starting with cassava and rubber. Not limited to them.
The control system is feedstock-agnostic by design. We are proving it out first on two feedstocks Thailand produces at scale — but the same sense-predict-adjust loop applies to any natural input whose batch-to-batch variability is the obstacle between it and a precision specification.
Where we are proving it
current material tracks
Wood and plywood adhesive
Structural bonding for panel manufacturers. The first market, and the one where bio-based chemistry is already cost-competitive.
EN 314-2 bond quality
No BPA in the backbone
No BPA in the backbone
Electrical insulation resin
Casting and impregnation resin for transformers and utility-grade equipment, where dielectric performance is the binding constraint.
IEC dielectric standards
Thermal class targets
Thermal class targets
Semiconductor sealing compound
Epoxy molding compound for chip packaging — the hardest specification we are working toward, and an active research track rather than a near-term product.
JEDEC / IPC reliability
Research stage
Research stage
EV and robotics seals
Sealing elastomers for drivetrains, battery enclosures, and actuators — the same control loop applied to an elastomer rather than a resin.
Thailand: world's largest
natural rubber producer
natural rubber producer
Bringing us your own feedstock
open to new inputs and specifications
A different input, or a different target spec
If you have a natural feedstock whose variability is blocking a precision application — agricultural residue, another crop-derived stream, a byproduct you currently downgrade — the system is built to be retrained rather than rebuilt. We are interested in these conversations.
Your feedstock
Your specification
Your specification
Team
Chief Executive Officer
Chief Technology Officer
Chief Scientist
Built by people who have shipped both halves of this.
Kawisorn Kamtue
AI PhD candidate at Carnegie Mellon University — physics-informed neural networks and industrial control. Former researcher on Apple's AI/ML Siri Understanding team; former data scientist at Tencent.
Jirameth Tarnsangpradit
PhD, Carnegie Mellon University — macromolecular engineering and polymer-grafted nanoparticles. Lecturer at KMITL.