Air Cargo Intelligence
Agentic Analytics over Indian Aviation Data
Captured from the live deployment, not a mockup
01 · The problem
Turned fragmented public aviation data into a product that answers ranking and anomaly questions directly, rather than handing users another raw data portal.
02 · How it works
- 01
Set measurable acceptance targets before building and reported against them: 99.6% of rows reconciled without manual mapping, 11.1% median forecast error, and every document either extracted or refused with a recorded reason.
- 02
Held alerts to 3.1 a month above a materiality bar, because an anomaly feed nobody can work through is the same as no feed at all.
- 03
Indian air-cargo data is spread across DGCA and AAI releases in incompatible shapes, so built an ingestion layer into a PostgreSQL star schema with an idempotent loader, reconciling 99.6% of rows (12,740 of 12,790) without manual mapping.
- 04
Competed five forecast candidates per series, naive, drift, recent-mean, seasonal-naive and SARIMA, picking the winner by rolling-origin backtest, reaching 11.1% median MAPE with 82.4% coverage on the 80% prediction interval.
03 · What it cost, and what it returned
Refused documents rather than guessing, recording a reason for each of the 15 refused out of 261, and held anomaly alerts to 3.1 a month above a materiality bar so the output stays actionable.