I design and ship AI systems, developer tools, and MCP infrastructure. On the research side I work on making AI efficient, auditable, and well-behaved.

I'm interested in AI that stays efficient, auditable, and reliable — keeping cost down while every answer remains checkable, and understanding how models behave at their edges. Representative published work:

• TACET (Latin: it is silent) — Cost-Amortised Reasoning via Self-Distilling Neuro-Symbolic Cascades: From Knowledge-Graph QA to Regulatory-Compliance Checking. A three-tier cascade — a sound Datalog engine, a calibrated ComplEx link predictor, and an LLM teacher — with an online distillation loop that mines teacher answers into auditable Horn rules, amortising LLM cost while every symbolic answer ships a replayable proof tree; the same mechanism carries from knowledge-graph QA to streaming GDPR compliance checking. code · preprint doi:10.5281/zenodo.20621240.

• Engineering Lessons — notes on the machine-learning systems I have built: the approach taken, how it lines up against industry practice, and what was still worth knowing once everything was running. Six posts so far, each in Vietnamese and English. source

/plugin marketplace add n24q02m/claude-plugins

Then /plugin install <name>@n24q02m-plugins. All 8 MCP servers plus the agent-chat plugin in one marketplace.

In beta, not yet on a package registry or in the marketplace: better-workspace-mcp — Google Workspace (Docs, Drive, Calendar, Gmail, Sheets, Slides, Tasks, Chat, People, Forms) with multi-account support.

Setup any server: Copy the Agent Setup guide link and send it to your AI agent with "Please set up this MCP server for me."

Shared principles across the MCP servers, built on the mcp-core foundation. The credential-handling ones (1, 6, 8) apply to every server that talks to an upstream account; better-godot-mcp is the exception -- it runs locally with no credentials, so it skips the relay and auth layers entirely.

1. Zero-Knowledge Relay -- E2E encryption (ECDH P-256 + AES-256-GCM). Relay server never sees plaintext credentials. URL fragment secrets stay client-side per RFC 3986.
2. Composite Tool Pattern -- One tool per domain with action dispatch: 4-17 tools per server instead of dozens of thin endpoints, saving LLM context tokens.
3. 3-Tier Token Optimization -- Compact descriptions (always loaded), help docs (on demand), and MCP resources (deep reference) keep the always-on tool schema small.
4. Tool Annotations -- readOnlyHint, destructiveHint, idempotentHint, openWorldHint metadata so the LLM knows tool behavior before calling.
5. Security Defense-in-Depth -- SSRF prevention, path-traversal containment, prompt-injection (XPIA) boundary tags around untrusted content, and error sanitization.
6. Multi-User HTTP Mode -- Stateless DCR (HMAC-SHA256), per-user session isolation, AES-256-GCM credential encryption at rest, OAuth 2.1 + PKCE S256.
7. Degraded Mode -- Server always starts, even without credentials. Help and config tools work. Data tools return setup instructions instead of crashing.
8. Zero-Config Relay Setup -- Auto-open browser, user enters credentials, server receives config via encrypted relay, saves to local config.enc.

• KnowledgePrism -- Chat with your knowledge. Ingest anything (URL, PDF, EPUB, DOCX, CBZ, audio, video, image, text) into a persistent per-project knowledge graph, then produce any format (translation, brief, podcast, slide deck, mindmap, quiz, export). One chat-first agent plans compound requests as a typed capability DAG; ProjectKG is the auto-maintained spine, not a feature. klprism.com
• Aiora -- A health and environment companion for daily wellness. getaiora.com
• LinguaSense -- Realtime knowledge transfer from any screen or microphone. Desktop + mobile. (coming)
• Akasha / GWM (exploratory research, not a shipped product) -- a graph-centric world model with an LLM teacher and continuous learning; early-stage.