Practical AI, shipped
end-to-&end.
I'm a software engineer and data scientist in Ho Chi Minh City. I train speech models and ship LLM-powered apps end-to-end. What follows is a survey of the work.
I began as a quantitative developer at Treehouse Finance, building the Go services that calculated real-time portfolio P&L and positions across EVM chains. The role was deliberately ambiguous — half quant, half engineer — and I learned to speak both languages. For three years I tuned the engine behind those calculations for throughput and correctness, wrote Solidity smart contracts for on-chain integrations, and translated between the research team and the build team.
In late 2024 I made a deliberate move to machine learning. I spent seven months at VNSilicon training Vietnamese and Thai text-to-speech models — sourcing speech data from the open web, processing it for alignment and quality, and running training across cloud GPUs. In parallel, on my own time, I built The Slime's Dictionary, an AI-powered English dictionary with built-in learning features.
What I care about is the gap between a research-grade model and a product a real person can use. Most of my recent work lives there: batched inference, voice cloning, content pipelines, clean APIs, frontends that ship.
The signal is in the tail.
Augurly is a public research site where I publish a model trained on OHLCV crypto data. The headline finding: at the highest-confidence tail of the model's output, precision lifts well above baseline — though recall, predictably, collapses.
The prediction task is framed with the triple-barrier method from financial-data labelling. For each bar, the model estimates whether ETH's price will cross an upper barrier — set at 4× the trailing 4-hour Average True Range — within the next four hours; a symmetric lower barrier is tracked the same way. Across eight years of ETH training data, the base rate for hitting either barrier sits at roughly 15%.
On the held-out test set, the high-confidence threshold (≥0.7) hits roughly 48% precision — a 3× lift over that ~15% base rate; the deeper tail (≥0.75) is sharper still. These are reference numbers from offline evaluation — they don't guarantee live performance, but they're an honest picture of where the model has predictive structure.
The site publishes the model's running price-and-probability traces live, so anyone reading it can watch the signal in motion rather than only trust a benchmark.
The Slime's Dictionary
Every word in the dictionary is illustrated with several real-world photographs pulled from Unsplash, so meaning is learned through visual cues rather than text alone. Around the dictionary I built spaced-repetition flashcards over the Oxford 5K, grammar lessons distilled directly from textbook source material, and cross-language definitions across English ↔ Vietnamese ↔ Thai — all content generated through LLM pipelines from book-quality sources. I designed and shipped every layer: backend, UI, infra, CI/CD.
dictionary.aloslime.com →arbitrary /ˈɑːbɪtrəri/
tùy ý, võ đoán — not based on reason or system.
resilient /rɪˈzɪl.i.ənt/
(adj.) able to recover quickly from difficulty.
The present perfect
Past action with present relevance. From Murphy §7.
Augurly
A public research site I built to publish a model's live price and probability traces on crypto markets. Test-set evaluation shows ~48% precision with a 3× lift at the high-confidence threshold — reference only, not a live-performance guarantee.
augurly.xyz →SparkTTS-VN
A fine-tune of Spark-TTS — an autoregressive TTS built on a Qwen2.5 backbone — over a Vietnamese speech corpus I curated and processed. The demo checkpoint runs on Hugging Face Spaces; intelligible read prompts, surprisingly natural prosody for an openly available stack.
live demo · HF Space →ZipVoice-VN
A compact 120M-parameter non-autoregressive, flow-matching TTS model trained on Vietnamese speech data during my tenure at VNSilicon. Small enough for near-real-time generation; competitive quality with much larger models on standard read prompts.
live demo · HF Space →BARK, up close
BARK is Suno's research-grade text-to-audio model. It isn't really a production model — too large to be practical at scale — but it's a strong proof that its approach to TTS (autoregressive language modelling over audio tokens) works. I refactored the codebase, added a batch-inference path from scratch (so the same model serves multiple requests in parallel), and trained a HuBERT semantic-token model so the system can clone voices from arbitrary reference audio. Wrapped in Gradio for hands-on evaluation.
live demo →source on github →– present
Independent AI Developer
Building The Slime's Dictionary solo, from architecture to ship. Productionizing open-source TTS (BARK batch inference, HuBERT voice cloning). Contributing fixes to LlamaIndex core.
– Mar 2026
Data Scientist
Trained Vietnamese and Thai TTS models (ZipVoice, SparkTTS variants). Sourced and processed large open-web speech corpora — alignment, filtering, normalization — for multi-language training runs on cloud GPUs.
– Oct 2024
Quantitative Developer · Backend SWE
High-throughput Go microservices for real-time DeFi portfolio P&L, positions and exposure across EVM chains. Schema design, query optimization, Solidity for on-chain integrations.
– 2022
B.Sc Data Science · 8.2 / 10
Graduated 8.2 / 10. Continued with Deep Learning & ML Specializations (Coursera), Quantization Fundamentals (DeepLearning.AI), Decentralized Finance Specialization. Native Vietnamese; fluent English (equivalent TOEIC 800+, three years collaborating daily with an English-speaking engineering team at Treehouse).
AI / ML
- PyTorch
- Transformers /HF
- LlamaIndex
- scikit-learn
- Pandas · NumPy
- Gradio · Streamlit
Backend
- Go /3y
- Python /FastAPI
- PostgreSQL
- MongoDB
- Microservices
- Solidity /EVM
Frontend
- Next.js
- React
- TypeScript
- Tailwind
Infra
- Docker
- GCP
- Cloudflare Workers
- Fly.io
- GreenNode GPUs
- CI/CD /Git
If you're building something AI-shaped, write.
I'm open to engineering roles — particularly AI engineering: building agentic systems, shipping LLM-powered products, or anywhere the model and the product around it both need to ship.
kzdmxq@gmail.com →