Tariful Islam Tarif

Built a production-grade LLM observability platform featuring prompt versioning, risk controls, and multi-model fallback routing (OpenAI → Ollama → Mock) for high availability. Designed automated LLM evaluation pipelines using ensemble LLM-as-judge scoring across quality, safety, and hallucination-detection dimensions. Implemented real-time trace logging, cost tracking, latency metrics, and drift monitoring to support continuous model governance. Developed role-based access control (RBAC), audit logging, and secure inference workflows for enterprise-grade compliance.

Engineered a low-latency ML inference API supporting multi-model and multi-version serving. Achieved sub-100ms inference latency using ONNX Runtime and Redis-backed intelligent caching, cutting redundant compute by 80%+. Implemented Prometheus metrics, health checks, and OpenTelemetry-based structured logging for full-stack observability. Deployed via Docker Compose and Kubernetes with autoscaling and canary rollout support for zero-downtime releases.

Built a multi-agent LLM system enabling natural-language querying of medical datasets (heart disease, cancer, diabetes). Implemented NL-to-SQL pipelines with automated routing between database query tools and medical web-search agents. Integrated LLM-based summarization to generate structured, evidence-based analytical insights for healthcare data.

Designed a computer vision pipeline for robust before/after image comparison under camera misalignment. Implemented ORB feature matching, homography alignment, and SSIM-based structural change detection. Generated annotated outputs, heatmaps, binary masks, and structured JSON summaries to support auditability.

Built a deterministic multi-agent orchestration system with strict turn control and debate sequencing logic. Implemented memory nodes, semantic duplicate detection, and replayable JSON-based event logging. Visualized multi-agent workflows as DAGs to support debugging and evaluation.