India Post pickup-message parser
Paste a message from the office's WhatsApp pickup group. This page mirrors the rules in
tools/parse_whatsapp_pickups.py — the Python file stays the source of truth; this is a demo of what it decides.
Record it would write
Why the rules stay (measured on 98 real messages)
| Approach | Agreement with the parser | Median per message |
|---|---|---|
| These rules | reference | instant |
| Laya decision model (zero-shot) | 56% | 1.69 s |
| Laya, Malayalam-script messages | routed to the multilingual checkpoint; Latin-script Manglish routes to english | — |
| Hosted LLM (free tier, single-word prompt) | 57.5% | 0.60 s |
| NVIDIA Nemotron-3-Ultra-550B | did not follow the one-word format | 0.40 s |
Every disagreement was hand-checked: on the pickup messages the rules were right and the models read them as chat/notice (e.g. “Jaaz boutique ready” → chat). The group writes about thirty recurring phrases in Manglish, which is what a keyword list is good at and a zero-shot model is not.