home / competition / datathon 2026
// prize distribution
📸 Add your prize distribution photos — drop them in images/datathon/ and create a manifest.json listing them. They'll appear here as a click-to-enlarge gallery.
// the question
2 in 5 children sitting the GP Contest scored below half their paper. The datathon asked: why? We structured the analysis as an elimination board — four candidate explanations, one key number each, cleared or convicted.
// four explanations tested
Hypothesis 1
Infrastructure
rho ≈ 0
cleared — no rank agreement with mastery
Hypothesis 2
Gender
tiny + flips sign
cleared — too small and inconsistent
Hypothesis 3
Place
<40% of variance
partial — most variation sits inside the same GP
Hypothesis 4
Specific skills
gap is skill-shaped
convicted — carry, borrow, word problems
// three findings that stuck
The regrouping cliff
On matched question pairs (same operation, same presentation format — one without carry/borrow, one with), a measurable share of children answered the plain question correctly then failed the exact same operation the moment it required a carry or borrow. These children don't need the operation retaught. They are missing one procedure — the narrowest teachable gap in the dataset.
The presentation penalty
The same arithmetic skill asked as an abstract number sentence versus wrapped in a Kannada word problem produces a sharp mastery drop — same paper, same children. This points at reading support, not more arithmetic drill. The remedy is not the one the curriculum usually reaches for.
Context-adjusted rank
Built two composite indices from the Karnataka At a Glance 2024-25 report — Infrastructure Access (roads + mobile) and Education Access (out-of-school rate + gender parity). OLS regression of district mastery on both (R²=0.56, 28 districts) surfaces which districts are doing more with less, and which are underperforming their own context.
// what we built
- EDA notebook — item-level mastery, variance decomposition (Division → District → Block → Cluster → GP), regrouping matched pairs, presentation penalty, skill prerequisite gate graph across all 9 papers.
- Index notebook — district Infrastructure and Education Access Indices from KAG 2024-25, OLS residuals, matched district twins (Belagavi/Chikkodi share every index value by construction — any mastery gap between them is unexplained by context).
- Synthetic dataset generator — 9 CSVs (3 years × 3 grades × 100 students) matching the real schema, for safe development without touching student data.
- 7-screen Streamlit dashboard — The Board → Telescope → Answer Sheet → Cohort of 100 → Ladder → Rank Shift → Twins. Built as a stepper: each screen is one beat in a single argument, not a filing cabinet.
- Data quality report — 7 districts flagged with explicit exclusion reasons: internet data plausibility failures, implausible zero out-of-school counts, ASER/KAG district boundary mismatch for Vijayanagara (split from Ballari in 2021).
// data sources
- GP Contest / Akshara primary assessment — item-level responses from 1.38M children, grades 4–6, 2022–23 to 2024–25, Karnataka government schools only.
- ASER 2024 Karnataka rural report — district-level reading and arithmetic outcomes, 30 districts.
- Karnataka At a Glance 2024-25 — 136 MB government statistical handbook: roads, mobile, internet, school enrolment, gender parity, population across 31 districts.
Python · pandas · scipy · NumPy · plotly · Streamlit · seaborn · matplotlib · OLS regression
what I learned
The most important finding wasn't in the data — it was the framing. A carry-and-borrow procedure gap is a cheaper intervention than rebuilding geography-based resource allocation. Good analysis doesn't just describe; it narrows the action space.
// back → competitions & awards