# Duy Tran Thanh — Staff AI Engineer > Staff AI Engineer and Ph.D. in AI with 9+ years across applied AI research, production GenAI, and enterprise software. Builds and leads multi-agent systems, RAG, LLM fine-tuning (SFT, GRPO, distillation), and LLM evaluation in regulated banking — shipped AI that passed 6/6 contractual KPIs at Techcombank and runs an LLM serving platform at 200M+ tokens/day. Based in Hanoi, Vietnam. Open to Staff AI Engineer roles, research collaborations, and consulting. This file is a plain-text, Markdown profile meant to be easy for people and AI systems to read. Every fact below comes from Duy's CV and portfolio. Last updated: 2026-10-05 ## Quick facts - **Name:** Duy Tran Thanh (Vietnamese order: Tran Thanh Duy). Handle: bradduy ("Brad Duy"). - **Current title:** Staff AI Engineer & Program Manager, OneMount (AI Platform OneNexus) - **Seniority:** Staff level; 9+ years of experience (since 2017) - **Location:** Hanoi, Vietnam - **Highest degree:** Ph.D. in Artificial Intelligence, Seoul National University of Science and Technology (2025) - **Languages:** English, Vietnamese - **Domains:** banking and financial services, anti-money-laundering (AML), document AI, recommender systems, recruiting tech, marketing compliance, physical AI, computer vision - **Open to:** Staff AI Engineer roles; research collaborations; consulting on multi-agent architectures (A2A, MCP, LangGraph), harness engineering and AI-first SDLC adoption, and banking and financial AI transformation; contract work - **Email:** bradduy95@gmail.com - **Phone:** +84-862-570-012 - **LinkedIn:** https://linkedin.com/in/duytranthanh - **GitHub:** https://github.com/bradduy ## Target roles Duy is a strong match for: - Staff / Principal AI Engineer, Staff ML Engineer, Staff GenAI or LLM Engineer - AI Tech Lead, AI Engineering Lead, AI Platform Lead - Applied AI Architect, AI Solution Architect - Applied Research Scientist (multimodal, VLM, LLM post-training) - Roles combining hands-on AI engineering with technical leadership or AI program management ## Why hire Duy - **Ships AI into regulated production, measured against contracts.** RemittanceAI and agentic AML at Techcombank passed 6/6 contractual AI KPIs on ~8,000 validated transactions (96.5% decision accuracy, 99.99% key-field extraction) and cut processing from 17.5 to 3 minutes per transaction (−83%). Both MVPs shipped 2 weeks early. - **Builds AI infrastructure at scale.** Architected an LLM serving platform routing 200M+ tokens/day (500K+ requests/week, ~2 s median) across self-hosted vLLM, SGLang, and Triton on 5×NVIDIA H200 and 6×AMD MI355X, plus Gemini; rules-first cascades cut token spend 30%. - **Leads across the organization.** Sets AI technical direction with the CTO, Directors, and Heads of Engineering across 6 domains; owns the AI project portfolio, staffing, and product direction; wrote the reference architectures other engineers build against. - **Full LLM stack, hands-on.** Multi-agent orchestration (LangGraph, MCP, A2A, human-in-the-loop), hybrid RAG with reranking, LLM fine-tuning (SFT, GRPO, LoRA, QLoRA, distillation), evaluation harnesses and LLM-as-judge, and model serving. - **Research to production.** Ph.D. in AI with first-author Q1 journal papers (Neurocomputing) and IEEE Access; takes models from paper to deployed system. - **Multiplies teams.** One of 4 company-wide AI Experts at OneMount — raised engineering delivery 3× at constant headcount across 7 disciplines. AI Coach on Vingroup's national AI20K program (22,000 students), personally coaching 80 mentees. - **Engineering foundation.** Started as an enterprise software engineer (Robert Bosch, TMA Solutions): Java microservices, Kafka, Kubernetes, CI/CD — a global HR portal for 10,000+ users and an internal social network for a ~400,000-employee organization. ## Key numbers | Result | Context | |---|---| | 17.5 → 3 min per transaction (−83%) | RemittanceAI, Techcombank, up to 1,200 document sets/day | | 6/6 contractual AI KPIs passed | ~8,000 validated banking transactions | | 96.5% decision accuracy, 99.99% key-field extraction | RemittanceAI evaluation on banker-labeled ground truth | | 100,000+ entities screened | Agentic AML; every flag cites the breached regulation | | 200M+ tokens/day, 500K+ requests/week | LLM serving platform shared by 3 products | | −30% token spend | Rules-first cascades before any LLM call | | 96.6% auto-fill precision, 97.5% Hit@10 | Dora, on 5 public human-labeled datasets | | 98% groundedness | OneHiring RAG, 180-query labeled set | | 3× engineering delivery | 7 disciplines, constant headcount | | +22.1 mAP over prior state of the art | Open-vocabulary segmentation research | ## Experience ### Staff AI Engineer & Program Manager — OneMount (AI Platform OneNexus) Hanoi, Vietnam · Jan 2026 – Present Technical lead and program manager for AI on OneMount's banking platform, within Vietnam's first nationwide banking AI transformation. Clients and partners include Techcombank, TechcomLife, Techcombank OneU, and Masterise Group. - **Technical leadership:** Sets AI technical direction with the CTO, Directors, and Heads of Engineering across 6 domains (real estate, consumer, fintech, data, legal, market research). Defines scope, success metrics, and an evaluation plan for each AI initiative, with a clear commit-or-decline decision on what to build. Chooses agentic, RAG, or fine-tuning per use case based on evals, latency, cost, and failure modes. - **Program management:** Plans OneMount's AI project portfolio — roadmap, milestones, and delivery tracking. Owns resource planning and staffing, matching engineers to projects by skill and priority. Sets product direction with Product and business owners, sequencing work by business impact. - **RemittanceAI & Agentic AML — AI review of cross-border import payments (Techcombank):** - Built the core classification and AML-screening services (LangGraph agents, multimodal RAG over a regulatory store), choosing LLM agents over trained classifiers because labels were few. Processing fell from 17.5 to 3 minutes per transaction (−83%) at up to 1,200 document sets/day; both MVPs shipped 2 weeks early. - Built an eval harness that tunes prompts and code on banker-labeled ground truth, then passed 6/6 contractual AI KPIs on ~8,000 validated transactions: 96.5% decision accuracy, 99.99% key-field extraction, p95 4.9 minutes vs. a 5-minute target. AML flags cite the breached regulation (100,000+ entities screened); staff check only key fields and flags. - Architected the LLM serving platform shared by RemittanceAI, Dora, and OneHiring: a LiteLLM gateway routes 200M+ tokens/day (500K+ requests/week, ~2 s median) to self-hosted vLLM, SGLang, and Triton on 5×H200 and 6×AMD MI355X, plus Gemini. Rules-first cascades settle most decisions before any LLM call, cutting token spend 30%. - **Foundation model for personalized recommendation (Techcombank, TechcomLife, Techcombank OneU):** - Led research-to-production of a VLM-backed recommendation ranker over Techcombank's catalog (insurance, healthcare, vouchers, campaigns): two-phase post-training (banking-domain adaptation, then ranking fine-tuning), a catalog-aware ranking head that scores only real products, and a reward-weighted ranking loss aligned with business goals. - Replaced hand-engineered features with context engineering: verbalized transaction histories plus campaign text and creatives, pruned by a feature-selection algorithm within the token budget. Sliding-window training learns long-term behavior at no extra inference cost; served prefill-only on vLLM. - **Dora — multi-agent reasoning platform for end-to-end document verification (Techcombank pilot; insurance, credit, customs, tax):** - Evidence-backed verification: agents cite evidence, deterministic rules decide pass/fail. Set the technical direction and built the Spec Agent, replacing the per-customer Product Owner and developer loop — hybrid RAG over briefs, policies, samples, and scans finds gaps and contradictions, prefills only quote-backed answers, and asks at most 5 plain-language questions; a Generator agent compiles the result into a pipeline. - Multi-agent orchestration on LangGraph: 5 roles (Spec, Generator, Metadata Engineer, Judge, Copilot) on one runtime, where a new role is configuration, not code. Plan-and-execute with plans checked by 8 deterministic laws, governed MCP tools, memory, human-in-the-loop interrupt/resume, and tenant isolation. - Benchmarked on 5 public human-labeled datasets: 96.6% auto-fill precision, 97.5% Hit@10 on Vietnamese legal retrieval, 100% of 34,824 spreadsheet rows read. SFT + GRPO distillation into a self-hosted model behind LiteLLM. Shipped on Kubernetes with 56 invariants and 3,300+ tests in CI; authored 10 of 30 architecture decision records (ADRs), aligning Product, Design, and the Java/Kotlin team on schema-first contracts. - **OneHiring & company AI agent platform:** - Owned the reference architecture for a no-code agentic studio on Dify (build-vs-buy: extend open source over in-house) — 4 agents across source-to-hire, sourcing autonomous, hiring decisions human-in-the-loop; RBAC/ACL by role, PII redaction, and a versioned MCP/A2A capability contract so new domains onboard without runtime changes. The blueprint 6 engineers build against. - Designed the RAG layer on Qdrant: hybrid retrieval (dense + BM25) with cross-encoder reranking, 98% groundedness on a 180-query labeled set, every answer cited; per-domain collections with tenant scope pre-filtered inside vector search (no cross-domain leakage) and hard-delete propagation for candidate erasure. - **Marketing Brand Brain (Masterise Group):** - Decomposed the CMO's marketing workflow into agent skills with typed I/O contracts, eval sets, and a human-oversight boundary per step. Made the build-vs-buy call on evidence by benchmarking self-hosted vLLM against vendor LLM/VLM APIs on accuracy, p95 latency, and total cost of ownership. - Designed the reference architecture: a governed policy knowledge base compiled into versioned executable rules, per-rule orchestration over multimodal RAG (policy text plus design-system images for VLM brand conformance), and an approval workflow with full audit logging. One model gateway owns routing, schema-validated output, and provider fallback, with prompt caching. Findings cite the rule version; a named approver signs off, with no autonomous publishing. ### Research Scientist — Physical AI Lab, Korea Evaluation Institute of Industrial Technology (KEIT) Seoul, South Korea · Jan 2025 – Jan 2026 - Advised hospital and healthcare partners on privacy-preserving federated learning (Substra) and embodied-AI adoption — robot-platform feasibility, perception and VLA stack selection, sim-to-real risk on ROS 2 / Isaac Sim — with a build-or-decline recommendation per use case. - Cut inference cost ~2× training-free via sparse mixture-of-experts (MoE) expert routing. Quantized PyTorch vision models to INT8/INT4 and served them with TensorRT on Jetson Orin NX (8× smaller, <1% accuracy loss), with custom CUDA kernels on the hot path. - Delivered 3× classification accuracy over CLIP on real-world agricultural imagery, and state-of-the-art open-vocabulary 3D scene understanding (+3% over the 2025 ICML baseline). ### Research Scientist — Machine Vision Lab, National Research Foundation of Korea (SeoulTech) Seoul, South Korea · Sep 2020 – Jan 2025 - Open-vocabulary segmentation: first-author research with vision-language foundation models — an automated knowledge-distillation pipeline on MMSegmentation where frozen VLM teachers generate multimodal pseudo-labels for zero-shot recognition of unseen classes with zero manual annotation; +22.1 mAP over prior state of the art. Multi-GPU training (DDP, mixed precision) with W&B tracking; real-time inference monitored for latency and drift. - Clothes-changing person re-identification: designed PAH-Net (Neurocomputing, Q1) — state of the art on three benchmarks (+7.3 rank-1 on PRCC cross-clothes); 107 FPS at 32.6M parameters, up to 8× fewer than competing generative models. Deployed on NVIDIA DeepStream with multi-camera RTSP ingest and Kafka streaming. - Real-time sports-broadcast ROI detector (MMDetection) exceeding contractual KPIs for a government-industry consortium; worked directly with the industry client on requirements. ### Senior Software Engineer — Robert Bosch GmbH Ho Chi Minh City, Vietnam · Sep 2018 – Aug 2020 - Designed enterprise B2B systems from scratch to production: service decomposition, REST/OpenAPI contracts, data modeling across Oracle and MongoDB, event-driven Kafka messaging, Java Spring Boot / Spring Cloud microservices. Owned authentication (JWT), load balancing, and Jenkins CI/CD for a global HR portal serving 10,000+ users at ~1,000 concurrent requests peak. - Architected Bosch's company-wide internal social network for a ~400,000-employee organization — horizontal scaling on Docker/Kubernetes, Redis caching, asynchronous Kafka. Presented technical proposals to German and Vietnamese stakeholders; core code reviewer. - Awarded Best Performance 3 times. ### Software Engineer — TMA Solutions Ho Chi Minh City, Vietnam · Jun 2017 – Aug 2018 - Full-cycle delivery (Angular + Java) of production client applications. ## Community leadership and mentorship ### AI Coach — AI20K in Action Program (QS Reimagine Education Awards 2026), Vingroup · VinUniversity Hanoi, Vietnam · Jul 2026 – Present - National program of 22,000 students; personally coaches 80 mentees with an architecture-first curriculum: requirement discovery, build-vs-buy and agentic-vs-deterministic trade-offs, non-functional design, and reviews that take prototypes to production with evals and quality gates. - Cohort 3 (Security): co-mentored with the VinSOC Head of AI Solutions on AI for cybersecurity — agents and guardrails for SOC and SIEM, and anti-money-laundering. - Cohort 4 (Banking & Fintech): main coach, guiding teams building multi-agent systems for tax, an AI secretary, CAD drawing extraction and cost estimation for civil engineering, and fraud detection for ride-hailing transactions. - Appointed Committee Member of AI in Action, Vingroup's AI community, setting the program's technical review bar. ### AI Expert Trainer — OneMount Hanoi, Vietnam · Jan 2026 – Present - One of 4 company-wide AI Experts. Trains engineering and business teams on prompt and context engineering and agent evals, and coaches applied AI engineers. - Raised engineering delivery 3× at constant headcount across 7 disciplines via agent harnesses, MCP integrations, and preset Claude, Gemini, and Codex configurations, with per-use-case token-cost models. ## Education - **Ph.D. in Artificial Intelligence** — Seoul National University of Science and Technology (SeoulTech), Seoul, South Korea, 2020 – 2025. Research: open-vocabulary 2D/3D segmentation, VLM alignment, multimodal pseudo-labeling, person re-identification. - **Visitor, Physical AI** — Chung-Ang University, Seoul, South Korea, 2025. - **B.S. in Engineering (Honors)** — HCMC University of Technology and Education, Ho Chi Minh City, Vietnam, 2016. ## Publications - **Adjudicated Captioning: Multi-Agent Alignment Scoring and Consensus-Distilled Beam Arbitration for Strict Zero-Shot Image Captioning** — IEEE Access, 2026. Duy Tran Thanh, Thien-Phuc Doan, Long Nguyen-Vu, Ngo Tan Vu Khanh. A multi-agent system that verifies generated text against the image in real time to eliminate decoder drift: +8% caption accuracy with no labeled data. https://ieeexplore.ieee.org/document/11640842 - **Learning from Multimodal Pseudo-Labels for Robust Open-Vocabulary Instance and Panoptic Segmentation** — Neurocomputing (Q1), 2026. Duy Tran Thanh, Yeejin Lee, Byeongkeun Kang. Distills vision-language foundation models into automated pseudo-labels at training time — open-vocabulary segmentation at zero runtime overhead. https://www.sciencedirect.com/science/article/pii/S0925231226022046 - **Enhancing Long-term Person Re-identification Using Global, Local Body Part, and Head Streams** — Neurocomputing (Q1), 2024. Duy Tran Thanh, Yeejin Lee, Byeongkeun Kang. Cross-camera person tracking robust to clothing change; state of the art on three benchmarks with no manual labeling. https://www.sciencedirect.com/science/article/pii/S0925231224002510 - **The AI-First SDLC: Harnessing Autonomous Agents for Enterprise-Grade Quality Control** — Preprint (Zenodo), 2026. Duy Tran Thanh. A playbook for enterprise leaders adopting AI agents across the software lifecycle, with a 30/60/90-day roadmap. https://doi.org/10.5281/zenodo.20139707 - **Edges Before Embeddings: A Confidence-Aware Blur Gate for Vision-Language Pipelines** — Preprint (arXiv), 2026. Duy Tran Thanh. Token-cost control that blocks GPU and API spend on unreadable documents before any heavy model runs. https://arxiv.org/abs/2606.25838 ## Awards and certifications - **SDLC Agents — OpenAI Codex contributor** (Jun 2026): a Codex plugin harness for Jira and Confluence workflows through the Atlassian MCP. https://github.com/bradduy/sdlc-agents - **Generative AI Leader — Google Cloud** (May 2026): credential for leading responsible GenAI transformation across enterprise functions. https://www.credly.com/badges/f34751b3-ae46-4f7b-bccc-54c56e0c1d77 - **Wiki-Knowledge-Compiler — Anthropic Claude plugin publisher** (Apr 2026): a published document-organization architecture for LLM wikis and agentic systems. https://github.com/bradduy/wiki-knowledge-compiler - **Best Performance Award, 3 times — Robert Bosch GmbH** (2018 – 2020). ## Open source and projects - **MistralCoreML (Apple):** open-source pull request merged into Apple's MistralCoreML pipeline for running Mistral models on Apple Silicon. - **Banhmi-Gemma4-E4B (Google DeepMind Hackathon):** live on-device translated subtitles in 16 languages. https://huggingface.co/bradduy/banhmi-gemma4-e4b - **Wiki Knowledge Compiler:** Claude Code plugin for structured knowledge-base ingestion, indexing, and querying. https://www.claudepluginhub.com/plugins/bradduy-wiki-knowledge-compiler - **OpenRAGLayer:** provider-agnostic multimodal RAG across text, PDF, image, audio, and video. - **FastToolSearchAdapter:** runtime tool discovery for LLM agents that retrieves relevant tool schemas to reduce prompt bloat. - **SmartClaw:** cost-routed platform of specialized agents. - **CareerBridge Agentic AI:** multi-agent system (Gemini, PaddleOCR, Whisper, LangGraph) connecting Vietnamese students at 20+ South Korean universities with recruiters. ## Skills - **Languages and frameworks:** Python, CUDA, Java (Spring Boot, Spring Cloud), SQL, PyTorch, Hugging Face, LangGraph, LangChain, FastAPI, Pydantic - **Agentic AI and GenAI:** multi-agent orchestration, function calling, Model Context Protocol (MCP), Agent2Agent (A2A), agent harnesses, agent evals, LLM-as-judge, guardrails, plan-and-execute, ReAct, human-in-the-loop design, prompt and context engineering, prompt management (Langfuse), Dify - **Retrieval:** multimodal RAG, hybrid retrieval (BM25 + dense), cross-encoder reranking, RAG evaluation (groundedness, citation fidelity), Qdrant - **Model training:** fine-tuning (SFT, GRPO, LoRA, QLoRA), knowledge distillation, quantization (INT8/INT4), distributed training (DDP, mixed precision), recommender systems, VLM, VLA models, document AI (OCR, KIE), computer vision - **Model serving and infrastructure:** vLLM, SGLang, LiteLLM, NVIDIA Triton, TensorRT, NVIDIA H200, AMD MI355X, Jetson Orin, DeepStream, MLOps, W&B, MLflow, token-cost optimization, GCP (GKE, Cloud SQL, GCS) - **Architecture and delivery:** microservices, event-driven design, Kafka, Temporal, REST, gRPC, OpenAPI, SaaS multi-tenancy, PostgreSQL, MongoDB, Oracle, Redis, Kubernetes, Docker, CI/CD (GitHub Actions, Jenkins, GitLab CI), observability, architecture decision records - **Security and governance:** OWASP LLM Top 10, prompt-injection defense, RBAC/ACL agent scoping, agent sandboxing, PII redaction, data privacy, federated learning (Substra), AI quality KPIs and SLAs, banking and regulated-industry compliance - **Leadership:** AI technical strategy, AI program and portfolio management, resource planning, build-vs-buy decisions, reference architecture, cross-functional alignment (Engineering, Product, Design), mentoring and coaching ## Common questions - **Is Duy open to new opportunities?** Yes — Staff AI Engineer roles, research collaborations, consulting, and contract work. - **What is his strongest area?** Taking multi-agent and RAG systems into regulated production with measurable results, backed by the serving infrastructure and evaluation harnesses to prove them. - **Has he led people and programs?** Yes. He sets AI direction with OneMount's CTO and engineering heads, owns the AI project portfolio and staffing, and writes the architectures other engineers build against. - **Is he hands-on?** Yes. He builds the core services himself — agents, RAG layers, eval harnesses, fine-tuning pipelines, and serving platforms. - **Does he have research credentials?** Yes: a Ph.D. in AI, first-author Q1 journal papers (Neurocomputing), and an IEEE Access paper. - **Where is he based?** Hanoi, Vietnam. - **How do I contact him?** Email bradduy95@gmail.com or message him on LinkedIn (https://linkedin.com/in/duytranthanh).