What IntelliForm actually does — results from the field
What IntelliForm actually does —
results from the field
Two platforms, real pilot data, and verifiable numbers. Here is what we measured when AI met the centrifuge floor — and what happens when you run a natural-language formulation brief through a five-model AI council.
I have spent the past year building two things in parallel: IntelliForm, an agentic AI platform for green chemistry formulation, and AI Node, a real-time process intelligence dashboard for centrifuge hardware at ChemRich Global's New Jersey facility. Both are now live, both have produced measurable results, and both are publicly documented on GitHub. This post is about the numbers.
I am writing this because there is a large gap between how AI in chemistry is discussed — generally, aspirationally, in market slide decks — and what it actually looks like when you run it on real equipment, against real regulatory frameworks, with real consequences for getting it wrong. I want to close that gap with data.
Part One: AI Node — What happened on the centrifuge floor
The AI Node dashboard (github.com/Cheme-Nova/ai-node) was built to do one thing: eliminate the fixed-timer mentality in centrifuge operation. Most industrial centrifuges run on predetermined cycle times — a chemist or engineer sets a duration based on experience, and the machine runs that duration regardless of what the product is actually doing. It is a blunt instrument for a precision problem.
AI Node uses Physics-Informed Neural Networks (PINNs) to detect the exact inflection point at which dewatering returns begin to diminish — the moment where additional spin time adds energy cost but not meaningful moisture reduction. Pair that with FFT vibration analysis that resolves bearing fault frequencies weeks before audible symptoms appear, and you have a fundamentally different approach to centrifuge management.
Here is what we measured over a 90-day pilot at a specialty resin facility, comparing the baseline fixed-timer operation against AI Node + ChemRich process control:
| Metric | Baseline | AI Node + ChemRich |
|---|---|---|
| Average Spin Time | 22 min | 16.5 min (−25%) |
| Moisture Variance | ±4.2% | ±0.6% |
| Energy per Batch | 14.5 kWh | 11.2 kWh (−23%) |
| Annual Capacity | Baseline | +18% |
| Downstream Rejection Rate | 12% | <1% |
The moisture variance result deserves specific attention. Going from ±4.2% to ±0.6% is not a marginal improvement — it is the difference between a product that passes downstream quality gates consistently and one that generates a 12% rejection rate. That 12% rejection figure represents real material, real energy, and real labor cost being discarded at the end of a production run. Bringing it below 1% restructures the economics of the entire batch.
"The inflection point detection replaces fixed timers — identifying the exact moment diminishing dewatering returns begin, triggering optimal discharge."
— AI Node README, github.com/Cheme-Nova/ai-nodeThe FFT bearing analysis component is less dramatic in its numbers but arguably more strategically valuable. Resolving BPFO, BPFI, and BSF fault frequencies weeks before symptoms appear converts unplanned downtime into scheduled maintenance. For a production centrifuge running continuous shifts, the difference between a planned bearing swap and an unplanned failure mid-batch can be measured in tens of thousands of dollars and several days of lost output.
AI Node is deployed on Streamlit Cloud at chemenova-ai-node.streamlit.app and the full source is MIT-licensed at github.com/Cheme-Nova/ai-node. The dashboard runs in real time against simulated centrifuge telemetry, with live FFT spectrum visualization, PINN-driven moisture prediction, and an ROI tracker.
Part Two: IntelliForm — From brief to certified blend in 30 seconds
IntelliForm (github.com/Cheme-Nova/IntelliForm, now at v0.6) is the formulation intelligence platform I have been building since ChemeNova's founding. The core claim — that you can go from a natural-language brief to a pilot-ready, certified, cost-optimized formulation in under 30 seconds — is specific enough to be falsifiable. Here is the architecture that makes it true, and the validation that backs it.
The optimizer: why ~8,000× matters
The formulation engine runs a surrogate-assisted NSGA-III algorithm — a multi-objective evolutionary optimizer — across cost, performance, sustainability, and safety simultaneously. The "~8,000× speedup" figure refers to the difference between running the optimizer directly against the full chemical property space versus using an ensemble of 18 XGBoost surrogate models to approximate the Pareto front first, then refining. This approach is the basis of the USPTO provisional patent currently pending on the platform.
The methodology is published. The full two-stage surrogate-assisted NSGA-III pipeline and EcoMetrics sustainability scoring framework appear in a ChemRxiv preprint (DOI: 10.26434/chemrxiv.15000857, 2026), validated against 47 real ISO 14040/14044 life cycle assessment datasets. Active journal submissions are in progress at JCIM and npj Computational Materials.
The QSAR engine: R² = 0.89
IntelliForm's QSAR engine uses Mordred's 1,613 molecular descriptors combined with a GradientBoostingRegressor to predict biodegradability, ecotoxicity, and performance properties from SMILES strings. The R² of 0.89 on biodegradability prediction is not a cherry-picked number — it is comparable to published benchmark performance on the OECD biodegradation dataset and grounded in the same methodological framework as Moriwaki et al. (2018). ChemBERTa-2, pre-trained on 77 million SMILES strings, handles domain-specific toxicology screening that no general-purpose language model can replicate.
Certification pre-screening: catching failures before they cost you
Every formulation IntelliForm generates is automatically pre-screened against eight certification standards: COSMOS, EPA Safer Choice, USDA BioPreferred, EU Ecolabel, Cradle to Cradle, OMRI, NSF/ANSI 305, and ISO 16128. The Certification Oracle produces pass probabilities, gap analysis, and a cost-benefit table before any formal lab testing is commissioned.
The economic case is straightforward. A certification failure discovered post-pilot — after lab time, reformulation, re-testing, and consultant fees — typically costs $50,000 or more. IntelliForm runs the pre-screen in three seconds. The platforms that do not offer this capability (spreadsheets, generic LLMs, contract R&D firms operating reactively) are leaving that $50,000 risk on the table for every formulation cycle.
On EU CBAM: The Carbon Border Adjustment Mechanism exited its transitional phase in January 2026. All specialty chemical imports into the EU now require verified carbon declarations. IntelliForm generates ISO 14067-compliant carbon passports — Scope 1/2/3 breakdown, CBAM-ready JSON export, ingredient-level attribution, blockchain audit hash — automatically, in two seconds. Consultants charge €3,000–€15,000 per product for equivalent studies.
The five-model council: no single-model trust
IntelliForm does not run on one AI model. It runs a council of five specialized systems, each handling the domain it was built for. Llama 3.3 via Groq parses the natural-language brief into molecular constraints in under 100 milliseconds. Claude Sonnet handles compliance reasoning across REACH, COSMOS, ICH, and CBAM documentation. ChemBERTa-2 runs the toxicology screen. Gemma2 runs the LP and multi-objective optimization locally — meaning sensitive formulation IP never leaves the user's environment. RDKit and MolT5 calculate molecular properties. GPT-4o arbitrates when model outputs conflict and aggregate confidence drops below 90%. The architecture is documented in the public repository.
The Negative Knowledge Graph: the asset that cannot be replicated
Every AI chemistry platform trains on what works. IntelliForm is the only platform that systematically captures what does not, and uses it. The Molecular Memory Network currently holds 847 documented ingredient incompatibility pairs — combinations that produce instability, discoloration, pH drift, or separation, captured from real pilot batch failures across seven verticals. It also holds 312 certification rejection patterns, each tagged with the specific regulatory gap that caused the failure.
This data cannot be published, scraped, or purchased. It exists only inside organizations that run the batches and record the failures. The methodology draws directly on Raccuglia et al. (Nature, 2016), the seminal work demonstrating that machine learning on reaction failures outperforms training on successes alone — extended here specifically to formulation stability and certification failures.
Why these results matter right now
The timing of what we built is not accidental. The AI-in-chemicals market is growing at a 32% CAGR toward $28 billion by 2034. Generative AI in chemicals specifically is projected at a 24.9% CAGR through 2035. These growth rates are driven not by hype but by two structural forces: regulatory complexity increasing faster than compliance team capacity, and the retirement of senior formulation chemists taking decades of institutional knowledge with them.
CBAM, REACH SVHC updates landing quarterly, COSMOS version updates, EPA Safer Choice expansions — the regulatory surface area is expanding continuously. The organizations that build AI-native regulatory intelligence into their formulation workflow now will not need to retrofit it later at three times the cost. That is the market condition IntelliForm was designed for.
The AI Node results matter for a parallel reason. Specialty chemical manufacturing in the U.S. is under margin pressure from every direction — energy costs, material price volatility, tightening quality specifications from pharmaceutical and food customers. A 25% reduction in spin time and an 18% capacity increase on existing hardware, without capital expenditure, is the kind of outcome that finances the next investment cycle. You do not need to buy a new centrifuge. You need to understand the one you have.
What we built it with
Both platforms are Python, MIT-licensed, and deployable from a single streamlit run app.py command. The full source for both is public on the ChemeNova GitHub organization.
Where things go from here
IntelliForm v0.6 is live now at intelliform.streamlit.app. The free version requires no account, no credit card, and no onboarding call. You describe what you need in plain English and receive a formulation with certification probabilities, a carbon passport, and supply chain substitution options — in under 30 seconds. Enterprise access, single-tenant deployment, and API integration are available via direct contact.
The next platform in the ChemeNova roadmap is PrintChem™, extending the IntelliForm core to additive manufacturing formulation — photopolymers, binders, and support materials for FDM, SLA, and SLS, with ISO 10993 biocompatibility screening for medical-grade 3D printing. Target: Q3 2026.
The manufacturing loop that connects IntelliForm digitally to ChemRich Global physically — where any formulation generated on the platform can be piloted as a real 50–200 kg batch in New Jersey, with a no-cost reformulation guarantee if certification criteria are not met — continues to be the integration that no software-only competitor in this space can replicate.
If you work in specialty chemicals, pharmaceutical formulation, personal care, or industrial process optimization and want to run a formulation, stress-test a certification scenario, or discuss what the platform looks like at enterprise scale, reach out directly: shehan@chemenova.com
Comments
Post a Comment