what i'm currently researching, and where it stands.

i'm not a researcher by training — this is me learning in public, digging into a real dataset from my own product, and being honest about how far along it actually is.

a longitudinal study of milk-procurement behaviour in a dairy cooperative, built on my own SaaS platform for dairy farmers.

in progress · exploratory analysis

goal: deeply understand farmer behaviour before modelling anything — then build a farmer-churn prediction system on top of it.

~2.2M daily records 6,693 farmers 8 societies ~1.6 years of data kerala, india
  • full data pipeline: firestore → clean, de-identified analysis panel.
  • entity resolution — recovered true farmer identities (4,428 → 6,693) via phone-based ids where depot-shared membership numbers collided. a novel data-cleaning contribution, not something i expected going in.
  • characterised data quality and distributions; established that SNF and rate are derived values, not measured ones.
  • found a seasonal (monsoon) supply rhythm. the bigger realization: the real challenge is retention, not production.
  • segmented farmers into 5 behavioural types; measured ~27% churn, detectable from first-90-day signals — low regularity, morning-only habit, low volume. notably, not milk quality.
next up: churn-prediction modelling using time-series representation learning.

the reading list behind the modelling work above.

self-supervised & time-series representation learning

BERT (Devlin et al.) ✅ completed
PatchTST ⏳ queued
TS2Vec ⏳ queued
TST (Zerveas et al.) ⏳ queued
CoST ⏳ queued
MOMENT ⏳ queued

time-series foundation models

Chronos ⏳ queued
TimesFM ⏳ queued
Moirai ⏳ queued

last updated: august 8, 2026