Use casesServices

Electricity loss and illegal-use inspection

Losses show on the transformer, not on the meter; inspection teams can visit only so many meters, and each finding needs photos, a report, and an accrual worked out by the regulation's method.

With the systemMeters ranked for a visit, field photos read, the file assembled, period and consumption computed; the specialist signs.

Who decides
Loss and illegal-use specialist

Field team · meter-change report · 4 November 2026

Meter-change report (EK-6) · meter M-4471

The meter was removed for laboratory examination and a sealed meter was fitted in its place; the terminal-cover seal was found broken, and the meter was photographed with its serial number visible.

→ photo 3 shows the broken seal · laboratory report confirms

Illustrative The document as it arrives. In red, what the system marks for your expert.

1. The work

Today

  1. SelectInspection teams are sent from tip-offs, high-loss feeders, and lists of meters whose consumption dropped; the list is long and the teams are few.
  2. ReadAfter a visit, a specialist opens the team's photos one by one to check the meter serial, the index, the seals, and the wiring against the illegal-use report (tutanak).
  3. AssembleConsumption history, earlier meter work, earlier findings, and the laboratory report are pulled by hand from billing, the meter-data system, and work orders.
  4. ComputeThe period and the consumption are worked out in a spreadsheet by the regulation's method for the case on the report, and a slip shows up only when the consumer objects.
  5. DecideThe loss and illegal-use specialist checks and signs the accrual; the invoice must reach the consumer within three business days of detection.

With the system

  1. SelectEach transformer's energy balance is computed in code, and the machine-learning model ranks the meters below it for a visit, each with its reasons; your planners choose the routes.Model
  2. ReadThe vision-language model reads the field team's photos: the meter serial and index, the seal numbers and their state, and any visible intervention in the wiring, and marks where a photo and the report differ.SLM
  3. AssembleThe agent assembles the file read-only: the report, the laboratory report, the photos as read, the consumption history, the last recorded meter work, and any earlier findings.Agentic AI
  4. ComputeYour written checks, run as code, compute the period and the consumption only where the regulation writes the method for that case; any other case goes to the specialist uncomputed.Your rule
  5. DecideThe loss and illegal-use specialist checks the file against the photos and the report and decides and signs the accrual, as before, under the regulation's procedure.Person decides

What lands on the desk

The specialist opens one file: why the meter was visited, what each photo shows and where it differs from the report, the laboratory result, the meter's history, and the period and consumption computed by the regulation's own method. Nothing is billed or cut by the system.

FIG. 1 · Inspection fileIllustrative
Residential meter · transformer T-114 · laboratory report receivedMeter intervention suspected, meter removed

Model · why this meter was visited

Top bandthis week's list for the region

  • positive: Consumption fell steeply and stayed low; similar homes on the feeder did not
  • positive: Cover-open events logged at night
  • positive: Loss on transformer T-114 rose in the same weeks

VLM · field photos

  • metMeter serial matches the reportphoto 1
  • metIndex matches the reportphoto 1
  • flaggedTerminal-cover seal brokenphoto 3
  • flaggedScratches on the cover screws, describedphoto 4

Agent · the file

  • ReportMeter intervention suspected; meter removed and replacedmeter-change report
  • LaboratoryIntervention in the meter foundlaboratory report
  • Last workSeal check, recorded with a reportwork orders
  • EarlierNo earlier finding on this consumeraccrual records

Code · the regulation's method

  • metPeriod from the last recorded meter work, within the limitart. 45
  • Consumption from the undisputed same periodart. 44computed
  • metConsumption already billed deductedart. 44
  • Model
  • SLM
  • Agent
  • Code

Why a person decidesThe accrual is the specialist's to sign under the regulation's procedure; the system never cuts supply, issues an invoice, or files a complaint.

DecidesLoss and illegal-use specialist

2. The system

For your technical team — skip it if you like.

How it works

In

  • Transformer meter readings
  • Smart-meter readings below it
  • Which meters hang on which transformer
  1. Code · calculates

    Transformer energy balance

    Energy into each transformer against the meters below it, in code

    Reads

    • Everything under In

    Out

    • Loss per transformer and period
    • Missing readings marked
  2. Model · scores

    Inspection-priority model

    Champion vs. challenger, validated out of time, calibrated

    Reads

    • Consumption and load profile per meter
    • Tamper, cover-open, and outage events
    • Loss per transformer
    • Past inspections and their outcome

    Out

    • A ranked list per region
    • The reasons for each meter
    • Likelihood calibrated on past outcomes
  3. VLM · reads field photos

    Field photo reader

    Fine-tuned on your photos of meters, seals, and connections

    Reads

    • Meter and terminal photos
    • Seal photos
    • Connection photos
    • The fields of the report

    Out

    • Meter serial and index read
    • Seal numbers and their state
    • Visible intervention, described
    • Where a photo and the report differ
  4. Agent · acts

    Inspection-file agent

    Fixed workflow; read-only

    Reads

    • The report and the laboratory report
    • Billing and meter-data history

    Out

    • The case the report names
    • Consumption history and readings
    • Last recorded meter work
    • Earlier findings on the consumer
  5. Code · calculates

    Period and consumption

    The regulation's written method, as code; no model

    Reads

    • The case on the report
    • Last recorded meter work
    • Undisputed same-period consumption

    Out

    • Period within the regulation's limit
    • Consumption by the written method
    • Consumption already billed deducted
    • No written method: to the specialist
  6. Decision

    Loss and illegal-use specialist

    Reads

    • The file, the photos, and the report

    Out

    • Decides and signs the accrual
    • The report stays the field team's
    • Objections are handled by your unit
How the agent's work is arrangedFixed workflow

Fixed workflow

Every finding needs the same file in the same order (report, laboratory report, photos, history, earlier work), so a fixed sequence is enough and easy to audit.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Inspection-priority model (machine learning)

  1. Target

    Illegal use confirmed by a signed report and not overturned on objection, per meter and time window.

  2. Data

    12–36 months of smart-meter readings and events, transformer balances, network data, and past inspections with their outcomes.

  3. Method choice

    Champion vs. challenger: a gradient-boosted model against a logistic baseline and the rules you use today; no method presumed better.

  4. Validation

    Out-of-time validation, calibration by rank band, and stability across regions and seasons; documentation for your own review.

  5. Monitoring

    New meter types, network changes, and seasons trigger a re-check; a set of visits chosen at random measures what the list misses.

Vision-language model (VLM) for field photos

  1. Test set

    300–600 field photos labeled by your senior inspectors before training: meter serial, index, seal number and state, visible intervention.

  2. Dataset

    Past field photos with their reports and laboratory results; people in the photos cropped or masked.

  3. Continued pre-training (CPT)Skipped

    Skipped: reading meters and seals needs no field language the base model lacks.

  4. Supervised fine-tuning (SFT)

    On photo and field pairs from your own meter types and seals.

  5. Reinforcement learning (RL)Skipped

    Skipped: reading fields is taught by fine-tuning, and no further reward is needed.

Where the agentic system stops

The field team writes and signs the report at the site; the loss and illegal-use specialist decides and signs the accrual under the regulation's procedure, and your planners choose which meters are visited.

Show its tools, limits, and record
Tools and access
Reads the report and photos from the field app, billing, the meter-data system, work orders, and earlier accrual records, read-only; writes only the inspection file.
May not
Decide that use was illegal, sign or change a report, issue an invoice or a payment notice, cut or reconnect supply, file a criminal complaint, contact the consumer, or grant itself any access.
Who decides and the written limit
Loss and illegal-use specialistThe field team writes and signs the report at the site; the loss and illegal-use specialist decides and signs the accrual under the regulation's procedure, and your planners choose which meters are visited. The system never decides that use was illegal, never cuts or reconnects supply, never issues an invoice, and never files a criminal complaint.
Stops when
A photo it cannot read with confidence, a photo that differs from the report, a missing laboratory report where one is required, or a case without a written method: the file goes to the specialist, marked.
Record
Every record and photo read, the model and rule versions, each computation and the article behind it, and who signed the accrual.

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. More confirmed findings per inspection visit.
  2. Accrual files complete and computed by the written method, so fewer objections upheld.
  3. Specialist time per file.

What your institution measures

  • Confirmed findings per inspection visit
  • Non-technical loss per transformer
  • Accruals overturned on objection
  • Time from detection to invoice
  • Specialist time per file

Measured in the Decision Blueprint

Today's confirmed findings per visit and accruals overturned on objection, from your inspection and objection records, set the baseline; your institution sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • The regulation names the cases of illegal use and the detection steps; where a meter intervention is suspected, the meter is removed and the illegal-use report (EK-5) follows only once a laboratory report confirms it (Elektrik Piyasası Tüketici Hizmetleri Yönetmeliği arts. 42 and 43). The field team and your unit do every step the regulation names.
  • Period and consumption follow the regulation's methods and limits for each case (arts. 44 and 45); the code applies only a method written there. The invoice goes out within three business days of detection, and the consumer may object within six months (art. 47).
  • Findings are photographed and, for the cases in art. 42/1 (a) to (c), also recorded on video (art. 43); the model reads photos only, and the video stays with the team and the unit.
  • The ranking learns from past inspections, which were not chosen at random, so a set of visits chosen at random measures what it misses. People in photos are cropped or masked, and meter data stay on your servers.

Start here

We take a year of inspections and their outcomes with your loss unit, compare today's selection with the transformer balances, and test the photo reader on your own field photos before anything is built.

Who builds it Every engagement is led by an engineer who has built machine-learning models that real decisions run on since 2012, and language models and agentic systems in production since 2024. About eqlx

Equalize the gap.

Bring work that takes expertise. Leave knowing the first step.

A short description of the work is enough; no confidential records.

  1. You send a short descriptionNo confidential records, credentials, or production data
  2. The engineer checks whether it fits, and what it would needA model, a small language model, an agent, or none
  3. A proposed first step, or an honest “not a fit”A technical conversation with the engineer

https://eqlx.ai/use-cases/illegal-use-inspection · Printed copy; the website is the current version.