Đinh Minh Trí — Senior Applied AI Engineer

01 — Experience

Where I've worked

Six years across enterprise AI, logistics platforms, and high-traffic products.

Feb 2026 — Present

Ho Chi Minh City

Senior Applied AI Engineer · Vinsmart Future — Vingroup

Enterprise Search — internal AI knowledge base

–Architected a scalable knowledge base on RAG, hybrid retrieval, vector search, and reranking — accelerating discovery across SharePoint, OneDrive, and internal repositories.

–Designed end-to-end AI pipelines: ingestion, dense embeddings, real-time query processing, access-control enforcement, and LLM observability with Langfuse.

Vivipedia — public AI encyclopedia

–Engineered a platform that autonomously drafts knowledge articles from official state news, with human-in-the-loop expert validation before publication.

–Built the multi-agent orchestration system for source synthesis, citation tracking, and fact-checking; scaled outputs to a reliable high-traffic production platform.

Clinical AI Suite — healthcare domain

–Patient Summary — production LLM service extracting structured clinical data (history, medications, allergies) from Vietnamese patient records; FastAPI on a hexagonal architecture with multi-agent section summarization (progression, problems, meds, vitals/labs, alerts).

–Medical Note Reconstruction — service that parses raw Vietnamese medical notes (outpatient, inpatient, emergency) into structured patient guidance: treatment instructions, home monitoring, nutrition and lifestyle, danger-sign warnings, and follow-up schedules.

–Established evaluation as an engineering discipline: golden datasets with evals running as a CI gate before every release, and every production error converted into a regression test case.

Aug 2025 — Feb 2026

Ho Chi Minh City

Senior Full Stack Engineer · The CrownX — Masan Group

Logistics Core Modernization

–Spearheaded migration of a nationwide platform from a legacy monolith to a decoupled, event-driven microservices architecture.

–Developed high-throughput streaming pipelines with Apache Kafka and Python/Node.js, processing thousands of supply-chain events per minute.

Supply Chain Visibility Suite

–Built RESTful APIs and responsive dashboards giving warehouse managers real-time visibility into inventory and nationwide fulfillment.

–Cut latency by 60%+ and eliminated sync bottlenecks through query optimization and scalable Azure infrastructure.

Sep 2022 — Aug 2025

Remote

Full Stack Developer · CHEK Technology

–VibeData Studio — led full-stack architecture for an AI-agent workspace using React, TypeScript, SSE, and the Claude Agent SDK to autonomously generate governed data pipelines on Microsoft Fabric.

–My Branding Loyalty Platform — engineered a backend handling millions of concurrent transactions on Django and Cassandra, with an SEO-optimized Next.js frontend.

–VPMilk Corporate Website — directed a team of 3 developers on Laravel, establishing rigorous CI/CD workflows and automated testing.

Mar 2023 — Sep 2023

Remote · Freelance

Full Stack Developer · Upwork

–Bilingual Dictionary Ecosystem — built a Vietnamese–English application driving a 70% improvement in vocabulary acquisition for active users.

–Optimized relational schemas and data-fetching strategies, improving retrieval speed by 50%.

Feb 2021 — Sep 2022

Remote · Freelance

Backend Developer · Upwork

–Fitness Enterprise CRM — built custom operational systems for regional fitness centers, automating admin tasks and lifting operational efficiency by 60%.

02 — Case Studies

Selected work, in depth

The architecture and decisions behind the two systems I'm most proud of.

Enterprise Search — RAG at enterprise scale

Vinsmart Future — Vingroup

An internal knowledge base indexing SharePoint Online, SharePoint On-Premises, and intranet systems into Qdrant, split into two planes: an offline ingestion plane (connectors, parsing, embedding) and a latency-sensitive online serving plane (retrieval, reranking, generation with citations).

Key decisions

–Hybrid retrieval over pure vector search — BM25 and dense retrieval run in parallel and fuse via Reciprocal Rank Fusion; exact-match document codes, out-of-domain Vietnamese vocabulary, and short queries all defeat embeddings alone.

–Permission-aware retrieval, never post-filtered — ACLs denormalized into every chunk and applied as query-time filters; user identity resolved through Graph API, LDAP tokenGroups, and SharePoint-native groups.

–Three-layer delete architecture — instant tombstone soft-delete, nightly full reconciliation, and a serving-layer ACL check as the last line of defense: no "ghost documents" ever linger in the index.

–Quality-score parser routing — primary parser with a score threshold, falling back to VLM-based parsing for complex scans and tables, so poor input never silently degrades answers.

QdrantAWS BedrockHatchetdltKubernetes (EKS)LangfuseGoPython

Clinical AI — structured data from Vietnamese medical records

Vinsmart Future — Vingroup · healthcare domain

A suite of LLM services turning raw Vietnamese clinical text into structured, validated data for hospital systems — patient-history summarization and medical-note reconstruction across outpatient, inpatient, and emergency visits.

Key decisions

–Hexagonal architecture — domain logic fully independent of FastAPI, the LLM gateway, Redis, and Langfuse; any adapter can be swapped without touching the core.

–Graceful degradation — Redis and Langfuse are optional dependencies; when unreachable, the service falls back to built-in prompts and keeps serving.

–PII protected by design — deterministic hash-token de-identification instead of simple masking; emails, phone numbers, and national IDs stripped from clinical text before any data reaches observability.

–Evaluation as a CI gate — golden datasets run before every release, and every production error becomes a regression test case.

FastAPIPythonRedisLangfuseDocker

03 — Open Source

Things I maintain

viparseMIT · PyPI

Vietnamese-first document loader for RAG pipelines. Converts legacy TCVN3/VNI/VISCII encodings to clean Unicode NFC, with diacritic-aware OCR for scanned PDFs and a commercial hosted API on top. Encoding detection uses a syllable-inventory scoring model with confidence thresholds, backed by round-trip verification tests and a corpus of real corrupted files.

PythonOCRVietnamese NLPRAG

04 — Skills

Technical toolkit

AI & Machine Learning

Claude Agent SDKRAG ArchitecturesVector SearchMulti-Agent SystemsPrompt EngineeringLangfuse

Languages

GoPythonTypeScriptJavaScriptPHP

Frontend

ReactNext.jsTanStack QueryZustandTailwind CSSshadcn/ui

Backend

FastAPIDjangoNode.jsExpressLaravelGin / Fiber

Data & Databases

Apache KafkadbtMicrosoft FabricPostgreSQLCassandraMongoDB

Cloud & DevOps

Microsoft AzureDockerKubernetesCI/CD PipelinesNginx

2014 — 2018

Education

B.Eng. Electrical & Electronics Engineering

Ho Chi Minh City University of Technology and Education