IPO Allotment Engine

Bidding closes on an IPO.

Millions of applications land at once.

SEBI's T+3 rule leaves one night to process them.

And there is zero room for error.

Exchanges send the bids.

Banks hold the blocked money.

UPI confirms the mandates.

Depositories hold the shares.

The registrar, my client, sits in the middle of all of it.

Ingest. Bid files pulled from the exchanges over SFTP.

Validate. Every PAN and demat checked in linear time. Duplicate PAN bids drop out.

Reconcile. Bids × bank × UPI, joined three ways in memory, across all cores.

Allot. Category rules first, then the lottery.

Hand back. Results returned to the exchanges, every number verified.

4× oversubscribed: 120 applicants, 30 lots.

Each application ID's digits are reversed to shuffle fairly. No dice anyone can nudge.

Now wipe it, and run it again.

Same seed, same 30 winners. An auditor can replay every draw.

10 million records.

Reconciled in 6.34 seconds.

About 1.6 million a second, with Python's GIL out of the way.

And 80% less lock contention on the database.

Infrastructure. OpenTofu, Step Functions, AWS Batch on Fargate scaling to zero, static NAT IPs for exchanges.

Database. Aurora PostgreSQL, Flyway, concurrent indexes, 4-byte enums, JSONB audit archives.

Data engine. Polars reconciliation, RE2 validation, PAN de-dup, external merge sort, the seeded solver.

Control plane. Go (Gin) API, pgx, Cognito + JWT roles, stage gates, an immutable audit ledger.