AI product development in Bangladesh: plan the system, not just the prompt.

The tiers below focus on LLM products: model APIs, retrieval-augmented generation, agents, MCP and tool integrations, evaluation, guardrails, observability, and the conventional web or SaaS application around them. They do not claim custom foundation-model training, computer-vision research, or a general data-science practice.

The delivery baseline includes 90 days of defect correction and launch stabilization; continuing evaluation, prompt changes, and new workflows are separate.

Good early candidates have repeated work, accessible examples, a reviewer who can judge quality, and a way to limit harm. Document search, classification assistance, drafted responses, structured extraction, internal research, and workflow triage can be tested with representative cases before a broad product is built.

Avoid measuring a prototype only through impressive demonstrations. Demonstrations are selected; production receives ambiguous, incomplete, adversarial, and unfamiliar inputs. Build an evaluation set from normal cases, important edge cases, known failures, and examples that require refusal or human review.

A focused proof can begin through the AI product engineering service . For broader budgeting and supplier evaluation, pair this guide with the general software cost guide and the software-company hiring guide .

A retrieval pipeline may include document inventory, parsing, cleanup, metadata, chunking, embeddings, search, re-ranking, context assembly, citations, and an update process. Each stage can remove or distort information. Representative retrieval tests should be separated from answer-quality tests so the team knows whether a failure came from search or generation.

Permissions must be enforced before private content reaches the model. Filtering a citation after generation is not a security boundary. For multi-tenant or role-based products, source access, caches, logs, traces, and exports must respect the same data boundaries as the conventional application.

Start with task-specific criteria: correctness, completeness, groundedness, format compliance, tool choice, refusal, tone, or reviewer time saved. Not every criterion can be automated. Use deterministic checks where possible and structured human review where judgment is required.

For high-impact workflows, define a human approval step or safe fallback. The right boundary depends on whether the output drafts text, recommends a decision, changes a record, sends communication, moves money, or controls another system. The ability to act should never exceed the product’s permission and review model.

MCP can provide a consistent way to expose tools and context, but it does not decide whether a user is allowed to perform an action. The application must map authenticated users and tenants to allowed tools, resources, parameters, and approval requirements.

Treat model-generated tool arguments as untrusted input. Validate them against server-side rules, limit data access, protect secrets, and require confirmation for consequential actions. Log decisions and results at a level that supports debugging and review without storing sensitive content unnecessarily.

Bangladesh teams may pay global providers in foreign currency, so model choice and consumption can affect both architecture and commercial operations. The product should track usage by feature, tenant, or customer where those costs matter to pricing.

Minimize sensitive data, separate tenants, redact where appropriate, and avoid placing secrets in prompts or tool context. Access control, audit logs, deletion, backups, incident response, and data export remain application responsibilities.

A Bangladesh-based delivery team may still use model, vector, observability, or cloud providers outside Bangladesh. The buyer should approve what data may leave its environment, provider retention settings, account ownership, billing currency, service regions, deletion behavior, and the fallback path before production data is connected.

High-stakes legal, medical, financial, employment, safety, or regulated use requires specialist review beyond ordinary product engineering. The project scope should identify the responsible experts and independent assurance rather than implying that model integration creates compliance.