EaseMail: Agentic Email Automation
A 10-stage LangGraph agent that reads customer-support emails and attachments, finds the right template among hundreds, and drafts a personalised reply that is quality-checked against the customer’s booking.
Built at EaseMyTrip. The interface shown is an illustrative mock-up; templates and customer data are not shown.
The problem
Support teams answer thousands of booking-related emails across 21 departments. Choosing the right reply from 434 live templates and filling it with the correct booking details by hand is slow and error-prone.
Customer emails also contain personal data and attachments such as e-tickets and scanned documents, so the system had to be accurate and privacy-safe as well as fast.
How it works
The architecture, stage by stage. Click any stage to see why it exists.
A hybrid extraction pipeline reads native PDFs with PyMuPDF and uses RapidOCR as a fallback for scanned documents and images, so e-tickets and attachments are understood as well as the email body.
Key engineering decisions
Graph, not a chain
LangGraph made conditional routing, retries and partial re-runs explicit. Each of the 10 stages can be tested and observed on its own.
Retrieval before generation
Grounding drafts in approved templates keeps tone and policy consistent, and the LLM personalises instead of inventing.
Confidence-gated output
Drafts must pass an 85/100 quality threshold, and failing drafts get targeted feedback instead of a blind retry.
Privacy by default
Reversible PII masking means the model never sees raw personal data, but the customer still receives a fully personalised reply.
Hard problems along the way
ChallengeRegenerating whole drafts for small tweaks wasted tokens and time.
FixRun metadata and feedback loops are persisted, so drafts can be recreated and refined without repeating earlier stages.
ChallengeScanned attachments broke text-only extraction.
FixA hybrid path uses PyMuPDF first and falls back to OCR, chosen per document.
Impact
EaseMail drafts booking-aware replies across every support department, and cuts redundant API calls and LLM processing cost along the way.
- Support turnaround time
- −30%
- Live templates covered
- 434
- Departments
- 21
- Quality gate threshold
- 85/100