LinkedIn · Portfolio · Medium · Hugging Face ·
Dhaka, Bangladesh 🇧🇩
I build AI systems that help people move from messy business data → useful decisions.
My current work combines conversational agents through multi-step Agentic workflows, self-correcting SQL, automated sandboxed analysis using agents, and demand-forecasting services on Google Cloud. I work across architecture, implementation, integration, and client-facing requirements discovery.
| Area | Systems I work on |
|---|---|
| 🤖 Agentic AI | Supervisor agents, specialist sub-agents, tool calling, guided workflows, approval gates, and stateful sessions |
| 💬 LLM products | Forecast Q&A, self-correcting SQL, product knowledge retrieval, grounded answers, and uncertainty-aware clarification |
| 📈 Forecasting | BigQuery ML time-series pipelines, back-testing, holdout scorecards, and promotion optimization |
| 🛠️ Production engineering | Async streaming, retries, idempotency, tenant-scoped state, reconnect-safe conversations, and failure recovery |
Turns a complex retail-planning workflow into a guided conversation: users can configure projects, launch forecasts, ask questions about results, and explore insights without exporting data or navigating every workflow manually.
An end-to-end BigQuery ML pipeline covering data preparation, feature engineering, back-testing, model training, forecast generation, enrichment, and promotion scenarios.
Structure-aware retrieval over legal and court documents with metadata, reranking, weak-result rewriting, citation-or-decline behavior, and clarifying questions when the system is uncertain.
LangGraph · Claude · Gemini · Vertex AI · BigQuery ML · Polars · Cloud Run · Pub/Sub
Useful AI needs more than a model call. I care about clear autonomy boundaries, measurable evaluation, cost and latency control, reliable tool use, observability, and graceful failure.
📫 Open to remote AI engineering opportunities · UTC+6