AI & Research
AI does the heavy lifting. An expert always signs off.
We use AI openly, for the work it is good at. The rule is simple: AI proposes, a person verifies and approves.
Trust, but verify.
What AI does
- SDS import. From a PDF, a scan or an image: all 16 sections into structured fields, with uncertain fields flagged for review.
- Substance recognition. Substances are recognised during import and linked to the substance database.
- Draft translations of free text. Descriptions, your own phrases and anything else that has no official translation.
- Checking existing translations. Accuracy, completeness and consistency, with doubtful translations flagged.
- Suggestions from your own data. Based on your own phrases, translations and substances.
What AI does NOT do (and never will)
- Does not guess classification. Classification is calculated from the composition by CLP/GHS rules, deterministically: the same composition always gives the same result, and that is why we guarantee it. We do not guess classification.
- Does not translate regulated text. H and P statements and other regulated phrases use the official translations from the legislation of each country, region and regulation, without AI.
- Does not use your data for training or for other customers. Your data works only for you. The models we train ourselves learn from publicly available safety data sheets published by manufacturers and suppliers, and from public chemical databases such as the ECHA C&L Inventory and PubChem, never from what you have given us. The external AI providers we work with do not use your data for training either.
- Does not approve or sign off a safety data sheet. A person always does that.
- Does not send documents to customers or authorities without human approval.
- Does not replace the chemical adviser. It saves them time.
Human in the loop: trust, but verify
- • AI drafts, a person decides: AI extracts data and drafts translations, and you review the SDS before it goes out.
- • You choose how to translate: your own phrases, professional (human) translators or machine translation. AI is one of the options, not an obligation.
- • Every change and approval is recorded: who, what and when.
Chemplora for your AI agents (Chemplora CLI)
The Chemplora CLI lets the AI agents you already use (Claude Code, OpenAI Codex, OpenCode and others) work with Chemplora natively: search substances and the register of 300,000+ entries, read and prepare products and safety data sheets, manage companies, work with translation memories and follow the change history.
The agent works with your account and your permissions, so you decide which commands it may run.
Chemical AI Research
Our research is carried out by Vuk Pavlović and Ivan Milenković. Papers are in preparation, and links will be added here as they are published.
Area What it means for you Status
When AI says "I don't know" The system recognises when it is not sure and hands the work to a person instead of guessing. We measure where the right hand-over point is Research
Risk-weighted accuracy Not every error is the same: a wrong supplier phone number and a missed hazard statement (for example H225) do not weigh the same. Accuracy is measured by safety risk Research
Chemplora domain language model A smaller language model built on the Chemplora domain that advises and suggests when working with SDS, classification and regulations. It too only suggests: classification stays rule-based and a person decides In production
Domain-model-driven SDS extraction A model of 527 fields, 30+ languages, validation and automatic correction of extracted data In production
Extraction per language Accuracy measured for each language; vision models, OCR+LLM and hybrids compared; local models Research
Multilingual chemical name recognition The same substance recognised regardless of the language or how its name is written; a benchmark built from PubChem and Wikidata Research
Contradictions within an SDS Detecting when a safety data sheet contradicts itself, for example between sections In production
ECHA PCN rule checks 83 implemented validation rules, checked against the ECHA catalogue In production
Analysis of a corpus of 500,000+ safety data sheets The basis for writing assistance and quality checks Being built
Identifier quality in SDS in the wild How many safety data sheets in circulation carry a wrong CAS number (bad check digit), an Index number entered as a CAS number, or an EC number that does not match the CAS number Research
Same product, recognised by composition Two safety data sheets are the same product if they share the composition, even when the name, language, manufacturer and revision differ Research
SDS versions across sources Reconstructing the chain of revisions, translations and redistributions of the same product Research
AI agents with measurable control When an agent makes a mistake, the system returns every violation with the reason and the correction, and commands wait for human approval. (See the Chemplora CLI section above.) Research
Classification differences between countries How much the classification of the same product differs from country to country Research
SDS quality by supplier and market What the quality of a safety data sheet depends on and which errors are the most common Research