[ Solution ]
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.
The panel knows the answer. You have to be able to ask.
Once the measurement layer is in place, data accumulates fast: minute-level power, hourly consumption, quality metrics, bill lines, alarm history. Most answers are already inside that data — the problem is that finding them means knowing which screen and which filter to open.
Enver removes that requirement. You ask in plain language; it looks at your site’s own data, answers, and writes underneath which data the answer rests on.
For the energy manager, that is a shortcut. For everyone who is not one — the shift supervisor, the technical director, the managing director — it is the door into the panel.
A real question, a real answer
Enver
connected to site data-
Why was consumption so high last Saturday?
- 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
-
Are we at risk of reactive charges this month?
- 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
Two layers: chat and audit
Chat is Enver’s visible face; half of its real work runs without you asking.
The chat layer is for free-form questions: “what changed versus last month”, “which were my five most expensive hours”, “what is that transformer’s night load”. Answers come from the site’s live data.
The audit layer is rule-based and runs continuously. For compressed air, 21 rules are defined: has specific energy left the catalogue band, has idle-running risen, is the set pressure needlessly high, is the leakage trend upward. Every rule produces a recommendation card, and three things stand openly on each card: which data, which assumption, estimated effect.
The separation is deliberate: free chat gives flexibility, but the audit is deterministic — the same data always produces the same finding. Outputs that touch money, like savings recommendations, come from the audit layer, never from chat.
Anatomy of a recommendation card
Every card the audit layer produces carries the same skeleton:
Set-pressure optimisation
audit rule R-07BASED ON
Over the last 30 days, discharge pressure averaged 7.9 bar; 6.1 bar was measured at the farthest point of use. ΔP 1.8 bar — above the band.
ASSUMPTION
1 bar reduction ≈ 7% compressor energy (conservative, lower bound of the literature). The threshold is adjustable per site.
ESTIMATED EFFECT
~5–7% compressor energy
Why this shape?
design principleA recommendation with a hidden assumption cannot be argued with — it gets applied blindly or rejected blindly. With the assumption in the open, the argument happens in the right place: “is 1 bar worth 7%, or 5% at our site?” That argument is productive; “is this AI making things up?” is not.
Thresholds can be overridden per site: if your machine or your operating regime differs, you change the rule’s limit — the rule engine uses your value instead of the default, and the card states which one is in force.
What Enver does not do
On an AI page, this is the section that actually earns trust.
It does not invent numbers. If the data needed for an answer does not exist, it says “we do not measure this”. A number without a source is more dangerous than a wrong number.
It does not control anything. Enver sits on the read side: it analyses, recommends, reminds. Stopping a machine or changing a setpoint are jobs that demand authorisation and an audit trail; they are not done from a chat box.
It does not carry data out. Your questions and your site’s data never feed another customer’s model. Analysis runs on your data, inside your account’s boundary.
It does not make the decision. A recommendation card does not say “do it”; it lays out the basis, the assumption and the estimated effect. The decision belongs to the engineer — Enver makes sure it is made quickly and well-informed.
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.
What it changes, and for whom
For the energy manager: the morning routine shrinks. Instead of touring eight screens, one question — “what was unusual overnight”. When deep analysis is needed, the screens are still there; Enver is the shortcut that takes you to the right one.
For the technical director: answers without having to learn the panel. Five questions before the monthly meeting, five sourced answers.
For management: the answer to “where are the savings” rests on data, not on a slide. Because every number Enver gives carries its source, the number in the board pack is defensible too.
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.