Profile Update Platform
A self-service, auditable way to request, collect, review and apply profile corrections — without ever overwriting what's already there.
- Role
- SDE 1 – Backend (Java)
- Company
- Standard Chartered Research and Technology (SOLV)
- Period
- Jan 2024 – Present
- Stack
- Java · DynamoDB · REST · Freshdesk integration
Context
Some active retailers have missing or incorrect business profiles: a missing GST document, a wrong address, an outdated shop name. Before this platform, corrections happened over phone and WhatsApp and were typed in by a support agent. Nothing recorded what was requested, who supplied it or who approved it. The credit team’s document backlog, which they tracked by hand in a spreadsheet, was the first use case.
The problem
Build one bounded, auditable, self-service flow. Any internal team can request corrections in bulk, the customer supplies them in-app, a reviewer approves or rejects each one with a reason, and approved changes land on the real profile without destroying existing data.
Design
- Five categories, treated uniformly: shop photos, business KYC, personal KYC (all image-based), business address and shop name (value-based). For value-based items a reviewer can approve, reject, or replace the value and approve that instead.
- Data model (DynamoDB).
- A request table with nested sections → items → images, and two GSIs: business + created time, and status + created time.
- An append-only audit table keyed by request, with a time-ordered sort key.
- Enums are stored as plain strings.
- One status function. Status is derived bottom-up (image → item → section → request) in a single function, which is the only place status logic lives. Items within a section are alternatives (any one approved document satisfies it). Sections are mandatory. The request has five statuses: awaiting submission, partially submitted, under review, completed, cancelled.
- One open request per shop, without locks. A new ingestion cancels any open request in the same transaction that creates the new one. “Latest upload wins” means races self-heal.
- Idempotent writes. Creation carries a creation-request id and each reviewer submit carries a decision-batch id, so retries are no-ops.
- Batch reviews in one transaction. Audit rows are grouped per (item, outcome) instead of per image. A 50-image review takes ~4–7 operations, well within DynamoDB’s 100-operation transaction limit.
- Non-destructive write-back, per section. As soon as a section is satisfied:
- new documents are inserted at the head of the proofs list, because invoicing reads the first GSTIN it finds, so head insertion keeps GST invoices correct
- an address correction creates a new address row and repoints to it
- nothing already stored is modified or removed
- Operational pieces.
- Bulk ingestion through the internal admin portal (250 rows per upload, keep-last de-duplication, batched lookups).
- A Credit-page banner with a 2-second timeout that fails closed, plus a kill-switch.
- Freshdesk ticket create/update on first submission.
- Signed image links that expire in 15 minutes.
- An alert each time a section’s rejection count crosses a multiple of 3.
Key decisions & trade-offs
Results
- A single engine any team can reuse for a new correction type without code changes.
- Every change carries a full audit trail: who requested, who submitted, who approved, and why something was rejected.
Tech used
- Java
- AWS DynamoDB (transactions, GSIs)
- REST
- Freshdesk API