Toan Viet D.

ABOUT ME I'm a Kaggle Competition Master and senior AI engineer with 8+ years of experience, currently building production autonomous agent systems for a US enterprise: agents that take over real operational workflows (incident management, engineering tickets, sales and tutoring) and run with limited human intervention. My expertise spans autonomous and multi-agent systems, agent platforms, agentic RAG, LLM evaluation and guardrails, and the cloud infrastructure to run them reliably and cheaply. HOW I WORK I ship at 10x speed with a Claude Code + Codex multi-agent engineering harness I built myself: planner, builder, repo-explorer, adversarial-reviewer and eval-judge subagents wired into plan → build → review → evaluate loops, with ADR gates and mechanical checks so throughput never costs control. This is how one engineer delivered an incident-management agent for 13 teams, a ticket-to-PR engineering pipeline, an agent platform's knowledge layer and five production agent migrations inside a year. I pair that with production-grade engineering: typed contracts, tests, clear module boundaries, observability from day one. I focus on systems that actually get trusted to act. EXPERTISE Autonomous Agents & Multi-Agent Systems I build agents that do work, not chatbots: a multi-tenant incident-management agent (~90 tools, versioned skills, LLM intent routing) that replaced manual triage, ticket routing, SLA tracking and status reporting; a five-agent pipeline that turns a Jira ticket into a merge-ready PR with human gates only at risk points; a multi-agent PR reviewer; reasoning tool-using sales and tutoring agents. Frameworks: LangGraph/DeepAgents, Agno, Hermes Agent, MCP (FastMCP, AWS Bedrock AgentCore Gateway), agent memory and skills architectures. Reliability, Guardrails & Evaluation My design rule: the LLM returns typed tool calls, deterministic code validates and executes. Server-bound tenancy and permissions, dedup and audit hooks on every tool call, compare-and-swap writes, per-job and per-tenant budgets, safe fallbacks, human-in-the-loop approvals, shadow-mode rollouts before any autonomy. Multi-turn eval harnesses (deterministic invariants + rubrics + LLM-judge ensembles) gate every release. Observability with Langfuse, OpenTelemetry and New Relic; all models behind a LiteLLM gateway so providers are swappable. Agent Platforms (No-Code Agent Builders) Core contributor to a config-driven agent platform on AWS Bedrock AgentCore: agent definitions as versioned data, generic multi-tenant runtimes, tools over MCP, so business teams change prompts, tools and knowledge themselves with no engineer and no deploy. Built its knowledge-base management plane, RBAC, serving-runtime warm pools (cold start 13 s → 2.4 s) and traffic-split A/B rollouts. Agentic RAG & Knowledge Ingestion Primary author of a multi-provider ingestion/RAG service behind 137 production knowledge bases: foundation-model document parsing (tables, charts, OCR), vision-LLM outlines, taxonomy tagging, hierarchical chunking, hybrid search + reranking, ~22 QPS at p95 0.4 s. 50+ RAG systems delivered overall, including Agentic RAG and GraphRAG (LightRAG). Vector stores: AWS OpenSearch, Qdrant, PgVector, Pinecone, Chroma. Sources: PDFs/Office files, images, video transcripts, CSV/Excel, MongoDB, websites, Google Drive, Notion, Confluence, SharePoint, email. Providers: OpenAI, Anthropic, Gemini, OpenRouter, AWS Bedrock; self-hosted open-weight models via vLLM, Hugging Face, llama.cpp. Cloud & MLOps Design, deploy and operate AI systems on AWS (ECS Fargate, Lambda, SQS, Bedrock, AgentCore, KMS/SSM), GCP, Azure, SageMaker and RunPod, with Terraform, Docker, Kubernetes and GitHub Actions CI/CD. Scalable, secure, cost-efficient: production agents at $0.007 per conversation. LLM Fine-tuning & Optimization Fine-tuned LLMs from 2B to 72B parameters across domains and languages; 3–6× inference speedups via quantization, distillation, TensorRT and vLLM. Main author of xTuring, a 2k+ star open-source LLM fine-tuning library (xTuring). Reach out if you want to move real work from people to dependable AI systems: I'll start with one well-defined workflow, prove reliability in shadow, and expand autonomy from there.