Afif Mansib
Chowdhury

Engineering challenge

A till that cannot trade is worse than a till that is slow. Every architectural decision — offline journaling, idempotent sync, hardware abstraction — follows from the fact that a network drop must never stop a sale.

Engineering challenge

Card taps were failing intermittently. Root cause was Android's foreground dispatch — it performs an intent round-trip and an NDEF check before handing over the tag, which is too slow for payment-speed reads. Replaced it with reader-mode polling and tuned IsoDep timeouts. Then added tag-presence debouncing, so a single physical tap cannot register as two reads — on a payment terminal that is the difference between one charge and two.

Engineering challenge

Cross-border latency was degrading transaction reporting. Rather than optimise the link, I inverted the dependency: writes journal locally first, then sync asynchronously as idempotent deltas with retry and jitter. A slow link now delays reporting freshness instead of blocking a sale — the failure mode moved from the customer to the dashboard.

Engineering challenge

The common failure is calling a printer library directly from the receipt screen and a card-reader SDK from checkout. It demos fine — until a supplier discontinues a model and the change touches forty files. Solving it once at the boundary means new device models certify without an application rebuild.

Engineering challenge

This was a critical-path dependency, not a translation exercise. Without it the Thai engineering team could not onboard, and the Thailand go-to-market could not start.

Engineering challenge

Inference cost is cost of goods sold when you resell agent labour. Routing classifies each call by difficulty and walks a priority chain — free-tier capacity first, paid models only when the task warrants it. Tenant credentials are Fernet-encrypted at rest and never cross a tenant boundary.

Engineering challenge

Models are good at reading and bad at counting. So the model reads and code counts — file parsing and all arithmetic happen in local tools, and a spreadsheet never reaches the model. Agents authenticate through the public API with a scoped key, exactly like an employee, so row-level security and approval gates stay enforced server-side. Nothing commits without a person.

Engineering challenge

Confidence is tracked at every stage, and each step degrades rather than drops the request: if validation fails the pipeline continues on the extracted specification, and if the database write fails the quote is still returned. A model failure downgrades the result instead of losing the job.

Engineering challenge

When the entire product promise is that the other side cannot see your number, "we were careful" is not a guarantee. The privacy boundary is an explicit response mapper: counterparty amounts and internal serial numbers are structurally incapable of appearing in the wrong response. Money is integer cents end to end, rate limits are tiered by abuse surface, and the e2e suite plays the full journey against a disposable database.

Engineering challenge

The hard part sits between fiat and chain: a payment must be verified, KYC cleared, token quantity computed, tokens issued and the ledger updated — atomically, with rollback on any failure. Built as an event-driven pipeline with idempotency keys at every step so a retried webhook can never double-issue.

Engineering challenge

The payment path came first, because campaign funding is only as good as it. Full lifecycle including IPN — the server-to-server callback that confirms payment independently of whether the user's browser ever returns to the success URL. Treating the redirect as confirmation is how you lose money on this gateway.

Engineering challenge

Double-booking was the risk. Unpaid bookings hold a slot via a MongoDB TTL index that releases automatically after thirty minutes, and confirmation only happens on the Stripe webhook — so an abandoned checkout frees the slot without manual cleanup.