Plain-language assistants are easy to build and easy to believe. On a smelting furnace, an unsourced answer is a risk. We think every answer should show its working.
On a night shift, when the furnace is behaving oddly and the senior operator is not on site, someone will want to ask a question. Why is power per tonne creeping up? What did we do last time slag got this thick? It is now easy to put a chat assistant in front of plant data and let people ask. The hard part is making sure the answer deserves to be acted on.
General-purpose language models are fluent. They will produce a reasonable-sounding explanation for almost anything. On a furnace, a plausible but wrong explanation is worse than no answer, because it can push someone to change reductant or power settings for the wrong reason. Fluency is not the same as knowledge of your furnace.
A new cook in a restaurant kitchen who asks “how long does the stock reduce?” can get two kinds of answer. One is a colleague guessing from memory. The other is the head chef pointing to the recipe book, the note from last winter when the stock came out thin, and the reason the time was changed. The second answer can be checked, and it teaches the cook why. We want the furnace to give the second kind.
In Furnace Twin, engineers ask in plain words and answers cite their sources. The answers draw on three kinds of material, and each is shown so the person reading can judge it:
| Source | What it contributes | What to check |
|---|---|---|
| Plant records | Heat logs, lab assays, electrical data and shift notes | Which heats, which dates, whether conditions were similar |
| Model runs | Estimates from the calibrated physics model | The range on the estimate and the model version |
| Literature | Your PDF library of papers and reports | Whether the reference applies to your furnace and ore |
Every answer cites its sources, and the model behind it is a local language model with citations, so plant knowledge does not have to leave your infrastructure.
The other half of citing is admitting gaps. If the records do not cover a situation and the model is outside what it was calibrated on, the honest answer is “we do not have evidence for this”. An assistant that always produces an answer is training people to stop checking. One that sometimes says “no record of this” is training them to trust the answers it does give.
Citations are only as good as what gets written down. That is why the product includes shift intake by voice in Hindi or English, which turns a guided spoken handover into structured data, along with heat and lab forms. The handover that used to be a notebook and a phone call becomes something the next question can cite.
For capturing what experienced operators know before they move on, see the operator knowledge capture playbook. For why a physics model is the right backbone for these answers, read physics-based vs data-only furnace models. The full set is on the Furnace Twin guides page. Process changes should still go through your plant’s approval and safety procedures.
Because a fluent but wrong answer can lead to the wrong change on the furnace. Citations let the reader check which records, model runs or literature the answer rests on.
Plant records such as heat logs, lab assays and shift notes, runs of the calibrated physics model, and the literature and PDF library.
Furnace Twin uses a local language model with citations, and like other Autowhat AI products can run on managed cloud, in your own VPC or on-premises.
See it on your own data. Furnace Twin — Ask the furnace, see the proof. Book a 30-minute working session with an engineer.
