AI Excellence Lead with 8 years building production AI and data platforms, now working at the frontier of agentic systems — where human–agent collaboration and agents operating directly on data platforms are reshaping how software is built. I look past trends to the non-obvious bets, identifying niches where recent advances become applicable, and hold a technical thesis under pressure until it ships. I lead by defining what to build and why, making the hard architectural calls, and staying in the code beside the team. Comfortable across research and delivery, from genotype dimensionality reduction to enterprise data platforms, with governance (least-privilege, on-behalf-of-user, GDPR/HIPAA) built into everything I ship.
| Leadership & Direction | Sets technical direction and architecture for teams up to 12; player-coach (decides and contributes code); influence without authority (reversed a McKinsey-designed system); method transfer across teams; executive-facing reporting |
|---|---|
| Agentic Systems | Cross-HAT collaboration; Context Highway platform (L0/L1/L2/L3 authority layers); agent-driven SDLC; agent governance (least-privilege, OBO) |
| AI & Data Platforms | End-to-end pipeline architecture; Azure Databricks; OpenSearch; Postgres / SQL Server; large-scale ingestion & transformation; sub-second query design |
| Engineering | Python (FastAPI, Dash, scipy, numba, PyTorch, ONNX); C/C++; R; JavaScript; git-native tooling; high-performance computing |
| Research & Theory | Algorithm engineering; complexity; ML; linear algebra, graph theory, numerical & combinatorial optimisation |
| Governance & Risk | Responsible AI; GDPR / HIPAA; model & data governance; privacy, compliance, audit |
| Languages | Vietnamese (native); English (fluent) |
.wiki/ across four authority layers; adopted by 3–4 teams.Design and deliver AI-powered systems and data platforms for enterprise clients; own end-to-end architecture and lead engineering teams. Work ~50/50 across direction-setting and hands-on contribution.
The vision. Enterprise software can be delivered an order of magnitude faster when developers and agents work as Human-Agent Teams (Cross-HAT) — if the organization standardizes how context and authority flow between people and agents, rather than forcing a single heavyweight framework. The leverage comes from human judgment plus a shared, git-native context: an operating-model bet, not a tooling bet.
First production proof — GreenCSR (Viettel). A corporate social responsibility platform for Viettel, a large Vietnamese payment provider. Delivered in 3 months against an original 300 man-month / 9-month estimate, at 36 man-months and $3,000 in token cost. Led 9 engineers with 1 sub SA, 5 full stack engineers, 1 tester, 1 BA and a PM.
Setting the direction. Rather than enforce tooling, I defined the invariants the team must share — architecture, work division, API contracts — and left each developer's agent workflow free. Before: "vibe requirements, vibe architecture," scattered documentation, each developer working sporadically with fragmented tooling (BMAD, Superpower, …).
.wiki/ carries shared context across four authority layers (L0 canon / L1 forum / L2 godview / L3 private workbench); facts climb only by explicit promotion.The vision. Covestro (FPT customer), a world-leading chemical manufacturer, ran on a McKinsey-designed, rule-based growth-planning system. The old engine produces suggestions with rejection rate of upto 80%.
The solution: I argued the data could predict within-customer growth better than hand-written rules, and set the direction to replace it with an ML-driven prediction engine — de-risking the change with no interface retraining and a better rationale attached to every suggestion so users could trust it without relearning. Our system relied on a modern Data Lakehouse architecture. I also lead the integration of a set of applications into an agentic platform system.
Result: 30% fewer rejected suggestions.
Designed and delivered scalable pipelines for AI/ML over real manufacturing telemetry: high-volume sensor ingestion with timestamp normalisation, sensor-drift handling, missing-data imputation, late-arriving events; plus a keyword×document matching architecture at sub-second latency. Delivered 3 production AI applications and a semantic search platform with full data-quality monitoring and alerting. Led 12 engineers.
Redesigned end-to-end data architecture for systematic, data-driven embedding-model evaluation: refactored ingestion, normalized schemas enabling new models without schema changes, and a judgment-collection pipeline over thousands of real queries. Production-grade with edge-case test coverage and schema/recovery documentation.
Voice of Employee (internal LLM sentiment/topic pipeline + leadership dashboards); Migration Pathfinder (secure on-prem AI fact-finding tool for cloud-migration assessments).
Developed core single-cell analysis methods — pseudo-time trajectory inference and bulk RNA-seq deconvolution (PyDWLS, under review) — and a microscopy cell-segmentation app (classical CV + semantic segmentation, backend architecture, performance optimisation).
Enhanced dimensionality-reduction methods for genotype data to surface latent biological factors.
Structural and functional analysis of protein families: implemented algorithms to detect sectors in sequence alignments; simulated statistical and topological structure to assess sector significance and stability. With Dr. Varoquaux. Surfaced new structural/functional patterns in protein evolution.
Co-developed rRice (R) and pyRice (Python) to automate genomic-data crawling and aggregation for Oryza sativa subsp. Japonica research.
Improved Extreme Learning Machine training via metaheuristic / swarm optimisation (PSO) for first-layer weights; validated on benchmark datasets.
Vietnamese naming: DANG is the family name.