Practical notes from the engineers who build agents, copilots and AI products.
Telehealth platforms need scheduling, video, prescriptions, payments and records, built to healthcare privacy rules. The hard part is not the video call but the consult lifecycle around it: identity, consent, the prescription record, refund rules, and an audit trail the regulator and the quality team both accept.
Audit architecture, dependencies, secrets, monitoring, backups and, for AI, prompts and evals before accepting responsibility. The audit takes one to two weeks, produces a written risk register and a day-one runbook, and tells both sides what the maintenance contract is actually covering before anyone signs it.
A shell app owns identity, wallet and navigation; verticals are independent modules on a shared microservices backend. The shell stays small, each module ships on its own schedule through a versioned contract, and identity, payments, notifications and orders are backend services every module uses, not rebuilds.
Indian SaaS needs Razorpay or similar for INR with GST invoices, Stripe for USD, and metering that feeds both. The design decision that matters is keeping one internal ledger of entitlements and usage, with payment providers and tax rules as adapters, so a change in either never touches the product.
Support agents answer admissions, fees and timetable questions in local languages on web and WhatsApp, escalating with context. A student support chatbot pays for itself in the admissions and fee-deadline peaks if it is grounded in the institution's own rules, connected to the student record and gated on every action.
Structured outputs and function calling force LLM responses into validated schemas so downstream code never parses prose. The model returns typed fields, the application validates them, and a failure is a countable error, not silent corruption. Here is how each mechanism works and the validation layer you still own.
Stream tokens for interactive tasks, queue long jobs with progress, and never block a request on a slow model call. Those three rules decide whether an LLM feature feels responsive or broken. This article covers the streaming stack, the job queue, progress and cancellation, and what the build takes.
Investors want real users, an AI capability with an eval score, usage analytics and a roadmap, not a slide deck with screenshots. A clever demo proves nothing now that anyone can build one in a weekend; evidence that real people use it, and that the AI part is measured, is what earns a second meeting.
Enterprise buyers require SSO, granular roles and audit logs; building them early is cheaper than losing the deal that needs them. Each one touches the data model, so retrofitting means rewriting authorisation checks across the product, while designing for them on day one costs a few weeks and a handful of decisions.