AI automation replaces repetitive manual work with software that uses language models, retrieval and tools. An AI agent decides which tools to call, such as searching your data or updating a record, to finish a defined task. Al Emran has shipped one in production, Smart Campaign at NO-BS (RAG, a vector database, tool and function calling over live product data), and builds agents and automations remotely from Dhaka.
AI automation is a fixed workflow where a language model handles one step, such as classifying a ticket or drafting a reply. An AI agent chooses its own steps: it reads the goal, decides which tools to call and checks its own result before finishing.
Most business wins start as plain automation. Agents are worth the extra complexity when the path to the result varies from case to case.
AI automation is worth building when the work is frequent, text-heavy or judgement-based and a person currently does it by hand. Research, triage, summarising, data entry from messy sources and first-draft writing fit well.
A plain script or rule is better when inputs are structured and the logic never changes, because a model adds cost and variance for no gain. Al Emran tells clients which side a task is on before building.
Retrieval-augmented generation (RAG) looks up the relevant passages from your own documents or database first, then asks a language model to answer from those passages. The answer can cite its sources and stays current without retraining a model.
RAG fits when answers must reflect your data: product knowledge, policies, customer records. It needs good chunking, a vector store, ideally combined with keyword search, and an evaluation set so quality is measured, not assumed.
An AI agent stays reliable in production through the system around the model: narrow tools with strict permissions, a test set that runs on every change, logging of failures, and budgets for cost and latency.
For actions with real consequences, such as sending a message or changing a record, Al Emran adds a human approval step until the agent has earned trust on real traffic.
The best business processes to automate with AI first are those where volume is high and a mistake is cheap to catch: lead research and scoring, support triage, internal search over company knowledge, report drafting and data entry from emails or documents.
Smart Campaign at NO-BS is an example. It replaced manual prospect research with an agent that discovers prospects, scores relevance and authority, and drafts personalised outreach.
Choosing an AI agent developer starts with one request: show an agent running in production on real data, how its quality is measured, how wrong answers are handled and how cost is kept under control. A demo is not evidence.
Also check data handling: where your data goes, what is logged, and whether the work can run inside your own cloud account.
Retrieval that answers from your documents and database, with sources, a vector store and an evaluation set.
Agents that take real actions in your product through well-defined tools, with permissions and an audit trail.
Prospect scoring, personalised outreach, triage and multi-step workflows, the kind of automation behind Smart Campaign.
Expose your product’s data and actions to AI clients through a secure Model Context Protocol server.
Embeddings and vector search that find what users mean, added to a product without a rewrite.
Test sets, failure logging, cost and latency budgets, and human-approval steps where stakes are high.
Pick one narrow task the agent must do well and define what good looks like.
Retrieval, tools and prompts on your real data, with an evaluation set from day one.
Guardrails, fallbacks, logging and cost controls before real users touch it.
Release behind a flag, watch real traffic and tighten the failures.
I am model-agnostic and pick per task on quality, latency and cost. For retrieval I use a vector store alongside your existing database, so you do not need a new platform to start.
Answers are grounded in retrieved sources, evaluated against a test set and constrained by tools with strict permissions. Sensitive actions can require human approval.
Yes. Most of my work is adding agents, search or automation to an existing product on its live data, behind a feature flag.
Both. A no-code workflow is fine for simple, low-risk automations. I build custom when you need reliability, your own data, tool permissions or lower running cost.
Tell me what you’re building, or the role you’re hiring for. I usually reply within one working day.
