An open monitor for the cold chain of milk. It reads a temperature log and says how likely it is that the milk has crossed its bacterial limit, and how sure it can be.
Explore theFor the Smart India Hackathon our team designed a cold-chain pod for dairy logistics: an ESP32, a DS18B20 temperature probe and a Peltier cooler, with an app that warns when milk has been too warm for too long. Building its shelf-life model raised a question the hardware could not answer.
N(t) = N₀ eμtThe probe records the temperature of the pack every few minutes, faithfully and cheaply.
How many bacteria the milk started with, which kind, how fast they wake up: no thermometer can see any of it.
When a monitor is wrong, is it the cheap sensor, the reading interval, the formula, or the hidden biology? We set out to measure each.
A complete, open system: real growth data, kinetic models, a monitor that turns a log into a verdict, a simulator that tests it, and three ways to run it.
√μ = b (T − Tmin)data/dairyguard.kinetics · growthlawsdairyguard.monitordairyguard.benchmarkcli · firmware/ · demo/$ pip install -e . && dairyguard assess examples/truck_fridge_failed.csv DairyGuard Kinetic 2026.10 | Pasteurised milk, psychrotrophic spoilage flora (literature parameters) Log: 288 readings over 72.0 h, max 26.0 C, 5.0 h above 8 C in 1 warm event(s) Shelf life used (textbook batch) 127 % [########################] Chance it crossed the limit 81 % [###################.....] Hours left if held at 4 C 0 h (typical batch) Hours left, cautious 0 h (9 batches in 10 reach this) Verdict: LIKELY OVER THE LIMIT Limit: FSSAI aerobic plate count limit for pasteurised milk, 3 x 10^4 CFU/mL
A C core keeps 64 plausible batches in under 2 kB on an ESP32, drives a green, amber or red light, and matches the Python monitor at every reading.
Six deliveries judged in real time, a trip you can draw with your finger, and the science behind every verdict.
Searched ComBase, the public database of bacterial growth experiments, for milk. Transcribed organism, temperature, pH, water activity, source study and growth rate for each record, and converted every rate from log₁₀ to natural-log units. Added a published fluctuating-temperature experiment (Lin et al. 2016) and the FSSAI limit.
Fitted five temperature laws by least squares on log growth rates, including a monotone neural network trained as an ensemble of five. Judged every law on data it had not seen: each Listeria study left out in turn.
Simulated deliveries minute by minute: fridges, warm spells, packaging, three thermometers, four reading intervals, and hidden biology drawn for every trip. The reference model decides the truth; the monitor sees only the log.
Scored every rule by AUC, compared rules on identical trips with a paired bootstrap, checked calibration, pitted the monitor against a model trained on 4,000 labelled trips, and reran everything to confirm identical results.
Showed that bacterial growth depends on a temperature history only through one running total, a biological clock, the same idea as Ozawa's generalized time in thermal analysis. It predicts every number in the loss budget.
Packaged the monitor as a Python library and command-line tool, wrote it again in C for microcontrollers, and built the browser app. Wrote it all up in a book and an architecture white paper.
We scored every way of reading a log on one question: can it tell the cartons that crossed the bacterial limit from those that didn't? A coin flip scores 50, a perfect method 100.
AUC = P(s⁺ > s⁻)
Real data. On studies they never saw, five laws including a neural network score within a narrow band; growth rates differ between studies by about 25%.
The monitor under test. As well calibrated as a model trained on 4,000 labelled trips, stable from 80 plausible batches, and robust to correlated biology.
"Crossing the limit" means the bacteria count passed the FSSAI hygiene limit for pasteurised milk. It is a count, not a taste test or a safety verdict. The heat test is a calorimetry measurement, simulated so far.
Milk at 4, 8 and 12 °C crosses the limit at very different times. Switch to the bacteria's own clock, which runs faster when it's warm, and the three curves become one.
That one running total is all a monitor needs from the log, which is why a cheap thermometer and hourly readings are enough, and why the biology is what's left.
dy/dτ = q/(1+q) · (1 − ey−ymax)Real growth rates differ between studies by a quarter. The biology a log can't see is real, and large.
A flexible neural law learns nothing new from growth rates alone. Better formulas aren't the answer.
The monitor's estimate of hours left is cautious but imprecise, because the batch is unknown.
Keeping DairyGuard Kinetic's proven core, the next model will learn each batch's biology from complete growth curves of milk under changing temperatures and simple tests at loading. Its blueprint is published before any lab data exist: the data we will collect, the architecture, the training plan, and the pass-or-fail tests it must meet. A first prototype already runs on simulated batches.
Package, command line, firmware and browser app; the book; the architecture white paper; the blueprint of the next model.
Five laws tested on 20 unseen Listeria studies; between-study spread measured; monitor against a trained model.
A pipeline for complete ComBase growth curves, and a lab protocol for milk under changing temperatures.
Plate counts of milk at steady and shifting temperatures, with simple batch tests on day one.
A journal paper, built on the lab results, once they stand.
Watch six deliveries live, draw your own trip, explore the evidence.
Open →
The whole study: story, models, data, proofs, every result.
Read the PDF →
How DairyGuard Kinetic is built: data, models, training, monitor, firmware.
Read the PDF →The next model, planned before the data: data, layers, training, pass-or-fail tests.
See the blueprint →Python package, C firmware, 27 tests, every number regenerated from one command.
GitHub →Incubators and plate counting for a few weeks, to measure milk at steady and changing temperatures.
Calorimetry labIsothermal microcalorimetry of milk samples, to test the heat test on the real thing.
Research groupsCounts in milk or chilled food under changing temperatures, with the log. Full credit given.
आहारशुद्धौसत्त्वशुद्धिः
āhāra-śuddhau sattva-śuddhiḥ
When food is pure, the mind is pure.
Chāndogya Upaniṣad 7.26.2