Sense
Capture gas-phase and liquid-phase readings from a handheld probe dipped into the milk
The situation
Amul, India's largest dairy cooperative, sources raw milk daily from a vast network of local dairy farmers through village-level collection centres. The cooperative model depends on farmers being paid fairly and quickly for what they deliver, and on the milk that enters the supply chain being genuine.
At the point of collection, the cooperative was working with:
- A reliable detection method — the milk culture test — that took several days to return a result
- Milk that had usually already been accepted, pooled, and processed by the time results came back
- No practical way to run lab-grade testing at hundreds of daily village-level collection points
- Pooled, quality-linked payouts, meaning a few adulterating farmers diluted the payout for everyone else
- No option to verify quality at the moment the milk actually arrived
As collection volumes grew and the farmer network widened, the cost of relying on slow, lab-based verification became harder to absorb. Every adulterated batch that went undetected diluted the payout pool and introduced quality risk further down the supply chain.
The challenge
The challenge wasn't accuracy — the lab test was already accurate. It was speed and reach: a solution had to work at hundreds of collection points, run by staff without lab training, and return a result before the milk moved further into the supply chain.
Our approach
We approached this as a sensing and inference problem rather than a lab-replacement problem — asking what could be measured instantly, at the point of collection, that would correlate reliably with adulteration.
Our objective was to:
Train
Train a classification model against lab-verified milk culture test results as ground truth
Estimate
Go beyond pass/fail to estimate the approximate degree of adulteration
Deploy
Pilot at Amul's Anand unit, then scale across the wider collection network
What we did
1
Designed a handheld probe for dip-and-read simplicity
The team built a handheld probe fitted with a gas and liquid sensor array. Dipped directly into a milk can at the collection point, it captures readings that shift in detectable ways when milk has been diluted or otherwise adulterated.
Kept deliberately simple:
- No sample extraction and no lab handling required
- No specialised technician needed at the collection point
- Dip the probe, wait, read the result
2
Trained a purpose-built AI model on lab-verified ground truth
Raw sensor output is noisy on its own. The model was trained specifically for this task, using the traditional milk culture test as ground truth — teaching it to recognise, from live sensor readings, the same chemical signatures the multi-day lab test was designed to detect.
The result is a model that classifies a sample as adulterated or genuine, and where adulteration is present, estimates roughly how diluted the milk is — all from a single real-time reading.
3
Validated the system where it would actually be used
The engagement moved through hardware design and sensor calibration, model training against lab-verified samples, and a pilot deployment at Amul's Anand milk processing unit — one of the cooperative's most established, highest-volume hubs.
Why Anand first:
- A high-volume, representative environment for real-world validation
- A deliberate choice to prove the sensor and model together before wider rollout
- A controlled setting to compare sensor readings directly against lab results
4
Scaled from pilot to a wider deployment
Following validation at Anand, the system moved beyond the pilot site to a broader deployment across Amul's collection network, with collection centre staff operating the device directly.
What made scaling possible:
- No lab infrastructure or trained chemists required at each site
- Collection centre staff need no training beyond operating the device itself
- A graduated concentration estimate, not just a blunt accept/reject decision
How it works
From a single dip of the probe to a collection-point decision — in about three minutes, with no lab involved.
1
Milk delivered by farmer
A farmer brings a can of raw milk to the village-level collection centre.
Village-level collection centre2
Gas & liquid sensor probe
A handheld probe is dipped into the can, capturing combined gas- and liquid-phase readings.
~3 minute reading3
Purpose-trained AI model
The model interprets raw sensor readings in real time, trained against lab-verified ground truth.
Trained on lab ground truth4
Adulteration reading
A genuine/adulterated classification, plus an approximate concentration estimate where relevant.
~96% accuracy5
Collection-point decision
Staff accept, flag, or query the delivery on the spot, before it is pooled with other milk.
Accept, flag, or query on the spot“
Trust is the currency our cooperative runs on — every farmer and every consumer needs to believe the system is fair. This gave us a way to verify that trust in real time, rather than days after the milk had already moved on.
Amul
“
I used to send a sample off and wait. Now I dip the probe, and within a few minutes I know exactly what I'm looking at. It's changed how I do my job every single day.
Amul
~96%
Detection accuracy
AI classification of adulterated vs. genuine milk
Days→~3 min
Turnaround time
Versus the traditional lab milk culture test
Included
Adulteration scale estimate
Approximate dilution level, not just pass/fail
Pilot→scaled
Deployment footprint
Anand unit, then wider collection network
The outcome
What changed at the point of collection:
- Verification time cut from several days to about three minutes, on-site, with no lab required
- The check shifted from a retrospective audit to a real-time gate before milk is accepted into pooled supply
- Testing became possible at collection points that previously had no practical way to verify milk quality on-site
- Collection centre staff can operate the sensor with no specialised lab training
- The model estimates the approximate degree of adulteration, giving the cooperative a graduated basis for handling borderline cases
- An unanticipated deterrent effect — awareness that fast, consistent testing was happening appears to have moderated attempted adulteration in the first place
Frequently asked questions (FAQs)
1) What is milk adulteration and why is it a problem for dairy cooperatives?
Milk adulteration typically means adding water or other substances to inflate the recorded volume of a delivery. In a cooperative with pooled, quality-linked payouts, it directly reduces what honest farmers are paid.
2) How does a handheld gas and liquid sensor detect milk adulteration?
The probe is dipped into the milk and captures a combination of gas-phase and liquid-phase readings that shift in detectable ways when milk has been diluted, feeding those readings to a trained classification model.
3) Why does the traditional milk culture test take several days?
It requires sending a sample to a lab and allowing it to develop before technicians can accurately read lactose and chemical concentrations — accurate, but too slow to act on before the milk enters the supply chain.
4) How accurate is AI-based adulteration detection?
In this deployment, the model reached approximately 96% classification accuracy distinguishing adulterated from genuine milk samples, validated against the traditional lab test as ground truth.
5) Can the sensor estimate the degree of adulteration, not just pass or fail?
Yes. The model estimates roughly how diluted a sample is, giving the cooperative a graduated basis for decisions rather than a single accept/reject threshold.
6) Does using this sensor require lab training for collection centre staff?
No. The device is designed for dip-and-read simplicity, and collection centre staff can operate it without specialised lab training.
7) Can a sensor-plus-AI approach be piloted before a full rollout?
Yes — this system was piloted at Amul's Anand milk processing unit to validate real-world performance before scaling to a wider deployment across the collection network.
8) How does Atomic Loops build AI systems for point-of-collection quality checks?
By treating the problem as sensing plus inference rather than trying to speed up an existing lab process — identifying what can be measured instantly and training a model against trusted lab ground truth.
Industry Services
Education
Focus areas - Education
© 2026 London School of DIGITAL BUSINESS.
All Rights Reserved