Three real runs. Unedited.
Three real runs.
Unedited.
After the parser and controller reliability upgrade, we captured three IntelliForm runs exactly as they came out — two successes and one honest infeasibility. Here they are, with full blend compositions, costs, eco scores, and regulatory output.
Demos are easy to cherry-pick. What I want to show here is the actual output of the upgraded IntelliForm controller — three runs chosen to represent the range of what the platform does today: a clean agricultural success, a food-vertical success with real regulatory flags, and a case where the optimizer correctly refused to produce a result rather than hallucinate one. All three were captured locally on April 24, 2026, after the parser, vertical-mapping, and constraint-propagation upgrades. Nothing has been edited.
LLM_PROVIDER=regex — the local regex-based parser rather than a live LLM endpoint. This is intentional: it isolates the optimizer's performance from any LLM variance, demonstrating that the constraint extraction, ingredient filtering, and optimization stack work correctly independent of the language model backend.
Run 1: OMRI Bioinsecticide Adjuvant
Agricultural vertical. The brief asked for an OMRI-listed bioinsecticide adjuvant for organic crops, under $8/kg. This is a constrained problem: organic compliance limits the available ingredient pool, and adjuvant performance requirements are specific.
| Canola Methyl Ester | |
| Ethanol (agri grade) | |
| Sulfur Dust (agri) | |
| Urea (agri) | |
| Ammonium Sulfate (fertilizer) | |
| Calcium Ammonium Nitrate |
What this shows: The optimizer resolved the agricultural vertical correctly, filtered to a 132-ingredient pool, and returned a feasible blend at $1.62/kg — well under the $8/kg ceiling. The eco score of 87.7 (Grade A) reflects the high canola methyl ester content, a bio-derived carrier. Both ethanol and sulfur dust came back OMRI-flagged as listed.
Run 2: Clean-Label Plant Milk Emulsifier
Food vertical. The brief asked for a clean-label emulsifier blend for plant-based milk, GRAS compliant, under $4/kg. The food vertical carries stricter allergen obligations than any other, and the optimizer surfaced them correctly rather than suppressing them.
| Water for Injection (WFI) | |
| Lauryl Glucoside (and) Glycerol | |
| Sodium Bicarbonate (agri) | |
| Resistant Starch · Psyllium Husk · Pea Fiber · Oat Fiber · Bamboo Fiber | |
| Pea Protein · Soy Protein · Whey Protein · Casein · Egg Albumen · Transglutaminase · Chlorella · Spirulina |
What this shows: The optimizer found a feasible food-vertical blend at $3.39/kg against a $4/kg ceiling — clean label compliant, 90.9% bio-based, eco grade A. But it also surfaced four mandatory allergen declarations without being prompted to look for them. A generic LLM would either miss these or note them as suggestions. IntelliForm flags them as mandatory, citing the specific regulatory framework (EU 1333/2008, Codex Alimentarius).
The performance score of 71.9 — slightly below the 75.0 constraint — is worth noting. The optimizer accepted this within tolerance rather than producing a false pass. The "Review Required" status reflects real gaps, not system failure.
Run 3: The honest failure
Industrial vertical. Low-VOC degreaser for heavy equipment — high flash point, strong grease lift, moderate foam, bio-based, under $4/kg. This is where IntelliForm does something most AI formulation tools won't: it refuses to produce a result.
- Increase max cost above $4/kg — high-flash-point bio solvents are inherently more expensive
- Reduce the bio-based floor below 80% to allow high-performance synthetic carriers
- Adjust max ingredient % to allow a dominant carrier solvent
- Switch to All Verticals mode to access cross-vertical high-flash solvents
Why this matters more than the two successes: The constraints here are genuinely in conflict. Bio-based solvents with high flash points and strong grease lift at under $4/kg do not exist in sufficient concentration within the current industrial ingredient pool to satisfy all three objectives simultaneously. After three relaxation rounds — each progressively loosening the constraints — the optimizer still found no feasible blend and halted.
An AI that halts and explains why is more useful than one that produces a confident answer to an impossible problem.
The output includes specific, actionable guidance on which constraint to relax and why — not a generic error message. It also completed the regulatory pre-screen on the pool itself, correctly returning no SVHC hits even though no blend was produced. The system continued to do useful work up to the point of infeasibility.
What these three runs actually prove
Constraints drive the optimizer
Cost ceilings, bio-based floors, and performance targets parsed from plain English actually propagate into the optimization — the $1.62/kg result against an $8/kg ceiling is not coincidence.
Vertical isolation works
Agricultural, food, and industrial runs each drew from separate ingredient pools (132, 217, 158 ingredients respectively) with no cross-vertical contamination.
Regulatory flags are real
The four allergen mandatory declarations in Run 2 were not prompted — they emerged from the regulatory layer against EU 1333/2008 and Codex Alimentarius automatically.
Honest infeasibility beats hallucination
Run 3 attempted three relaxation rounds before halting. It returned a diagnosis and four specific remediation paths — not a fabricated blend.
Eco scoring is independent
Both successful runs returned Grade A eco scores (87.7 and 83.4) computed independently of the optimization objective — a separate accounting, not a reward metric.
No LLM required for core function
All three runs used the regex parser backend. The optimization stack — constraint extraction, filtering, solving — runs fully locally, independent of any LLM provider.
Try it yourself
The platform is live at intelliform.streamlit.app. No account, no credit card. Write a formulation brief in plain English — application, cost target, bio-based requirement, certification needs — and see what the optimizer returns. The full source is MIT-licensed at github.com/Cheme-Nova/IntelliForm.
If your formulation needs move toward a physical pilot batch, ChemRich Global runs 50–200 kg pilot manufacturing from our New Jersey facility, with a no-cost reformulation guarantee if certification criteria aren't met. Contact: shehan@chemenova.com
Comments
Post a Comment