Practical notes from the engineers who build agents, copilots and AI products.
Self-serve onboarding with KYC, catalogue and payouts is what lets supply grow without operations headcount. Treat the merchant onboarding flow as a product with an owner and a measured funnel: verify identity and bank details through APIs, tolerate imperfect catalogue data, and show partners what is blocking approval.
Track active users, tasks completed in-system, workarounds still in spreadsheets and support tickets at 30, 60 and 90 days. Those four software adoption metrics, read by role and by site, tell you whether the platform is doing the work it was bought for, and each review should end with three fixes rather than a slide.
MCP is a standard for exposing tools to AI agents with scoped permissions, so agents can act on your systems without direct database access. You write one MCP server per system, define each tool's inputs, outputs and scope, and every agent, whichever model runs it, calls the same contract under the same audit.
Sellers use AI for listing copy, price monitoring, review responses and support across marketplaces from one system. The value is in the "one system" part: a catalogue of record that pushes listings to each marketplace, pricing rules with floors the seller owns, and a support agent that sees orders on every channel.
Marketplaces succeed on supply onboarding, payments and settlement, trust and disputes; the storefront is the easy part. Budget most of the build for the seller side, the ledger and operations tools, and treat the buyer app as the thin layer it is. Here is what the plumbing contains, what it costs and who builds it.
Logistics ERP centres on order intake, dispatch, fleet, proof of delivery, cash reconciliation and seller settlements. The build order matters: get order intake and dispatch working with a driver app first, because proof of delivery and cash data from the field are what make reconciliation and settlements possible.
Minimum LLMOps is tracing, an eval suite on every change, cost dashboards, model pinning and a rollback for prompts. A team of two or three engineers can run all five with open-source tooling and a few days of setup, and without them an LLM application in production is being operated on hope rather than evidence.
Use ML for prediction, classification and detection on structured data; use LLMs for language; many systems need both. The LLM vs machine learning choice comes down to the shape of the input, the cost of a wrong answer, the volume of decisions and the budget per decision; most production systems route between the two.
Observability traces each request with prompt version, model, latency and token cost so any bad answer can be located in minutes. It makes evals, routing and cost control auditable, and it is a few days of work with tools such as Langfuse or OpenTelemetry. Here is what to record and where teams go wrong.