Get in touch Panel login

AI Assistant: Enver

Instead of hunting for the right chart, ask the question directly: "Why was Saturday's consumption high?" Enver looks at the data and answers with its sources. An answer it cannot source, it does not give.

Get in touch

Chat

A real question, a real answer

Enver is connected to site data: hourly consumption, the site calendar, alarm history. The answer arrives with its source; it does not say “probably”, it names the data it looked at.

Chat history belongs to the site; ask the same question a month later and the answer is produced from fresh data with a link to the earlier one. Enver also explains alarms: clicking “TX-2 analyser communication lost” tells you at what hour and from which gateway it originated.

  • Natural-language question, sourced answer
  • No personal data; the record stays with the site
  • On the same screen as the panel

Enver first gathers the data behind a question with its own queries, then interprets it; beneath the answer it lists which tables were used. Only the question and summary figures leave the site, never raw measurements or personal data. Where it cannot answer, it says so rather than inventing.

Enver

connected to site data
  1. Why was consumption so high last Saturday?

  2. Enver

    Saturday's total was 26,412 kWh, 18% above the average of the previous four Saturdays. Most of the difference comes from the compressed-air line: the compressor room drew 3,804 kWh that day against a usual Saturday level of ~2,100 kWh. Of the compressors, unit C-132 stayed on load all day; the production log shows no Saturday shift. A compressor on load with no production usually means a line valve left open or leakage rising, I would check the per-line flow screen.

    • hourly consumption · 4 weeks
    • compressor power data
    • operating schedule
  3. Are we at risk of reactive charges this month?

  4. Enver

    Low right now. Mid-month, average power factor is 0.97 against the 0.95 chargeable threshold. But the last ten days trend downward: on the day shift it dips to 0.955. At this rate the threshold could be crossed towards month end. If you want the correction panel's stage check added to the maintenance list, I can set a reminder.

    • period reactive registers
    • hourly power factor · 10 days
The line under each answer is not decoration: Enver shows the data sources every answer used. No data, no answer, it does not produce guesses.

Recommendations

From measurement to recommendation, assumptions in the open

A rule engine scans the data with 21 rules; every recommendation card states the saving, the assumption and a confidence grade. Thresholds can be overridden per site; an applied recommendation is tracked through verification.

Each rule looks for a measurement pattern: steady flow out of hours, a capacitor bank left in overnight, COP below the band. A finding becomes a card; the card states the basis of the saving calculation, the assumption and the confidence grade. An applied recommendation is tracked against the baseline in the verification module.

Once a recommendation is applied its card moves to “tracking”; the savings verification module takes the period before as the baseline and normalises the period after against production. The realised saving is written on the card; the gap between forecast and actual feeds the model’s own correction.

  • Thresholds can be overridden per site; defaults stay in code
  • Cards ranked by saving × confidence
  • Recommendations needing capital are flagged for grant eligibility

Recommendation Cards

this week · whole site
  • Close the receiver outlet valve out of hours

    9,248£/yr

    leak analysis 92 tracking
  • Step the set pressure down (7.3 → 7.0 bar)

    1,803£/yr

    ΔP measurement 81
  • Switch correction out overnight

    1,273£/yr

    reactive map 88
  • Take Chiller 2 in for maintenance

    2,050£/yr

    COP band 64

Annual total 14,374£/yr

Schematic: the confidence grade shows how far the assumptions are backed by measurement; a low-confidence item is verified on site first.

What it answers, and with what

Capability provenance

Free-form questions
chat
The site's live and historical measurement data
Rule audit
21 rules
Compressed air · continuous, deterministic
Recommendation card
data+assumption+effect
Audit layer · thresholds per site
Comparison
period/period
Same day type, same calendar correction
Source display
every answer
Data sets used are listed under the answer
Scope limit
read-only
Analyses; never controls equipment

The rule library grows module by module: compressed air’s 21 rules are live; cooling and generator rules are next. As each module lands, Enver’s audit reach grows with it, the chat layer already sees all measurement data.

Frequently asked

Which model does Enver use? Does our data go into model training?

Enver uses a large language model as the interface; the content of the answer comes from your site's database. Your data is never used in another customer's analysis and is not sent for model training. Questions and answers stay inside your account's boundary.

What happens if it answers wrongly?

Two safeguards. First, the source line: every answer states the data it rests on, so checking is one click away. Second, the layer separation: outputs that touch money, savings recommendations, deviation findings, come from the deterministic rule engine, not from free chat. The same data always produces the same finding; there is no 'creativity' there.

Can it send commands to equipment?

No, deliberately. Enver is read-only. Work that needs remote control (starting a generator test, say) is done in its own module, with an authorisation chain and an audit trail. Machine control from a chat interface would be a risk, not a convenience.

Which languages does it support?

Like the panel itself, Enver works in English and Turkish. Whichever language the question arrives in, the answer comes back in the same one.

Are the chat examples real?

The examples are representative, publishing a real customer conversation would mean publishing that site's data. But the behaviour is exact: the source line, refusing to answer without data, and the recommendation-card anatomy work in the panel precisely as shown.

Let us measure what is happening on your site.

In a one-hour call we look at your existing setup and set out exactly which measurement points are needed and what you would be able to see.

Get in touch