- 🔭 Physical AI & Robotics: Training Vision-Language-Action (VLA) models and Imitation Learning policies (ACT, Diffusion Policy) on physical hardware (SO-ARM-101) with synchronized multi-camera pipelines.
- 🧠 Agentic AI & Cognitive Memory: Building autonomous Multi-Agent Systems (MAS) with planning-first reasoning, dynamic skill routing, and long-term memory architectures (epistemic integrity, MCP & A2A protocols).
- ⚙️ High-Performance Data Architecture: Designing high-throughput data lakes (Databricks Delta Lake, Apache Spark) and semantic Knowledge Graphs (SPARQL, Jena Fuseki).
- 🏆 Grand Prize Winner: Microsoft Agent Hackathon (Tokyo Electron Device / Microsoft Enterprise Division) — Multi-sensor robot anomaly diagnostics achieving 0.95 F1-score.
- 🎓 Background: Computer Science graduate from Hanoi University of Science and Technology (HUST K67) with Distinction (CPA: 3.78/4.0); 2x National Mathematical Olympiad Prize Winner.
Only repositories that are actively maintained are featured here. Archived work is listed separately at the bottom.
| Repository | Description | Topics |
|---|---|---|
| my-meta-brain | Owner-controlled encrypted second brain for AI agents (.NET 10 / C# / WPF): owner-scoped MCP access over single-use tokens and named pipes, DPAPI + AES-256-GCM vault at rest, encrypted backup & restore, selection-scoped export, and content-schema migration with encrypted rollback points. Active development. | agent-memory mcp dotnet security |
| mosaic-slm | Research framework for compute-aware multi-agent decision-making on one 16 GB consumer GPU: shared small language model, per-agent private memory (SQLite), and a shared temporal knowledge graph (Kùzu). Phases 0–2 complete; Phase 3 memory implemented with 0.0% deterministic paired improvement (target ≥5% not met); serving, training and RL not started. | multi-agent slm temporal-kg research |
| Repository | Description | Topics |
|---|---|---|
| projecta | Local-first Windows desktop knowledge workspace (Tauri v2, bundled local services): ontology-backed project graph, source-grounded notes, human-in-the-loop review of AI-extracted candidates, and evidence-backed queries without external runtime dependencies. Projecta 1.0.0 released Oct 2026 (Windows 10/11 x64). | ontology knowledge-graph tauri-v2 local-first |
| Repository | Description | Topics |
|---|---|---|
| latent-anything | Python framework treating latent space as a first-class object: PCA/UMAP/SAE introspection, steering vectors and activation patching, world-model planning lanes, and bounded-memory streaming pipelines. v1.0.0 published on PyPI (Sep 2026); documented evidence is CPU/synthetic-workload scoped. | latent-space interpretability pypi python |
| anomaly-representation-learning | Self-supervised representation learning for hard multivariate time-series anomalies: masked prediction, EMA targets, contrastive objectives, and a reproducible benchmark harness. Experimental; no verified accuracy claim. | time-series anomaly-detection self-supervised pytorch |
| Repository | Description | Topics |
|---|---|---|
| aegis-f | Offline-first personal Android expense tracker (Kotlin, Room/SQLite, ONNX): exact whole-dong VND ledger, optional bank-notification review inbox where every bank-derived record needs explicit confirmation, and on-device category suggestions applied only with explicit user acceptance. In development; no public release. | android kotlin on-device-ml offline-first |
| daily-research | AI research publishing platform (Next.js 16, TypeScript, Tailwind v4, MDX): overview, weekly and deep-dive tiers with bidirectional cross-references, plus RSS and Microsoft Teams tooling. | nextjs mdx research typescript |
| coding-agent-stacks | Reusable rules, skills and subagent definitions for coding-agent workflows, shipped with a PowerShell installer. Configuration payload — not an agent runtime. | agentic-ai developer-tools powershell |
| triet4p.github.io | Personal portfolio site (React 19, TypeScript, Vite) publishing the projects and write-ups linked from this profile. | portfolio react typescript vite |
These repositories are archived and no longer maintained. They are kept as a record, not as current work.
Discontinued experiments & products (archived)
| Repository | What it was |
|---|---|
| cogmem | Cognitively grounded long-term memory API with typed memory networks and hybrid retrieval |
| arch-lens-ai | Desktop R&D operating system for technical due diligence (Tauri v2 + React 19) |
| paper-kg-ontology | Research-paper knowledge graph lab: OWL2 RL reasoning, TDB2, SHACL validation |
| ecommerce-agent-databricks | Support agent on Databricks Apps with Lakebase, MLflow ResponsesAgent and governed tools |
| llm-serving-lab | vLLM / Ollama / FastAPI serving benchmarks behind an OpenAI-compatible client |
| claw-code-explain | 12 agentic harness patterns — Claude Code internals explained |
| bridge-research | Desktop AI research assistant (Tauri + React + FastAPI) |
| math-ai-connect | Mathematics and AI learning journal (Astro + MDX) |
| cognition | Rust cognitively grounded agent memory system, extended into an agent architecture |
| aegis-ma | A2A (Agent-to-Agent) SDK prototype over JSON-RPC/WebSocket |
| itapia | Explainable AI co-pilot for retail investors |
| mcp-fin-test | Financial agent built on the Model Context Protocol |
| otto-rec-sys | Recommendation-system solution for the OTTO competition |
| ez_cli | Rust CLI utilities for learning and performance testing |
| excel2md | Layout-preserving Excel to Markdown converter for Windows |
| magma_obby | Flutter game deployed to iOS via Codemagic |
Earlier study & competition archives (archived)
| Repository | What it was |
|---|---|
| app-log-process-k8s | Big-data pipeline (Hadoop, Spark, Kafka, Hive, dbt) processing application logs |
| SeEvoDJSSP | Self-evolution algorithm applied to the dynamic job shop scheduling problem |
| KaggleCompetition | Collection of Kaggle competition solutions |
| SearchAppQt | Qt app visualizing linear and binary search algorithms |
| NoteApp | Simple desktop app for taking and sharing notes |
