Practical notes from the engineers who build agents, copilots and AI products.
Permission-aware retrieval enforces document ACLs at query time so users only receive content they are allowed to see. The entitlement check runs in the retrieval layer before ranking, never in the prompt, and is proven with negative tests where a named user must not receive a named document.
Load-test to peak, cache aggressively, queue writes and rehearse the day; the sale is the design load. Peak-season readiness means treating the busiest hour of the year as the number your architecture is built for, proving it with a rehearsal weeks before, and having a runbook for the moment something still breaks.
Design and load-test for month-end and festival peaks, not average days; queues, caching and autoscaling are non-negotiable. Indian consumer apps spike on salary days, sale events and match nights, so the architecture must absorb bursts, shed load gracefully and be rehearsed under test before the real day.
Integrate UPI intent and collect flows, cards and wallets through a gateway, with webhooks and reconciliation on the backend. The app collects the payment attempt and shows status; the backend creates orders, verifies webhooks, decides success and handles refunds. Getting that split right is most of the work.
Benchmark on your tasks; open-weight models often match on retrieval and extraction and lag on hard reasoning, which routing can cover. Public leaderboards tell you which model is best at leaderboards. A golden set of a few hundred of your own tasks, scored the same way for every candidate, tells you what to ship.
On-device models handle voice, vision and quick tasks offline; cloud models handle reasoning; route by latency and privacy. In practice: small models on the phone for wake words, transcription, OCR and classification, a cloud model for judgement, and a router deciding per request by network, battery and sensitivity.
Offline-first means a local database, queued actions replayed in order, and written conflict rules; it decides the data model, so start there. The app reads and writes locally, a sync engine pushes and pulls in the background, and every record type has a rule for when two edits collide. Retrofitting this is a rewrite.
Natural-language reporting maps a question onto a semantic layer, runs a validated read-only query and returns a chart with definitions. Without the semantic layer it is text-to-SQL guessing at what "active customer" means. This article covers the architecture, the guardrails, accuracy and the build.
Multilingual AI needs per-language benchmarks and code-switch handling; the same discipline that works for Indian languages works globally. Build evaluation sets from real conversations in each language, benchmark models per language, treat mixed-language turns as normal, and let native speakers judge before launch.