Analyst in mind. Builder by hand. Contributor at heart.
I work in equity research and build AI and data systems around problems I encounter in practice.
- Adaptive AI: when should a system retrieve more evidence, spend more compute, or stop?
- Evaluation: how do we test systems on realistic tasks and consequential errors?
- Research systems: how can models move beyond retrieval while preserving evidence and provenance?
- Interactive explanation: how can complex systems become easier to inspect and understand?
I learn mostly by building: reproduce → adapt → break → understand → rebuild.
📊 vnibb
Vietnam-first equity research platform combining financial data, research workflows, a multi-widget dashboard, and database MCP integration.
Domain-specific model adaptation for institutional financial research, built with Huy X. Dang. The experiment achieved an 85% evaluation win rate in the AutoScientist Challenge.
Experiment Dataset · Write-up · Fine-tuned Model
An Obsidian-native knowledge system combining notes, graph structure, and vectorless retrieval into a browsable research garden.
Reusable interactive explainers for making complex systems easier to inspect and understand.
More things I've built
-
🤖 model-finetune Fine-tune Qwen 3.5 and Gemma 4 on private data and serve GGUF models locally. Model
-
🧠 agent-swarm-pack Compact multi-agent workflow and orchestration experiments.
- ⚖️ V-Legal Vietnamese legal archive with vectorless retrieval, citation graphs, briefs, and cross-references.
Archived / deprecated experiments
-
📸 gapsnap Gap-analysis snapshots at a glance.
-
🔀 sankeydrawer Interactive Sankey flow-diagram builder and exporter.
-
⚡ deepseek-n8n-automate-workflow Self-hosted n8n + Open WebUI + Qdrant + Ollama workflow stack.
-
📝 n8n-template-and-documentation-for-RAG Production-oriented RAG workflow templates for n8n.