Bioinformatics · Data Science · mRNA × AI · cfDNA fragmentomics · Formal methods · Lean 4 + Python · HKT
Developer · formal-verification hobbyist · ML-for-biology tinkerer.
I'm Royce — a developer who splits my time between formal verification, applied ML for biology, and the occasional late-night PR review. I like problems where the math has to check out and the code has to ship.
🎓 Background: M.S. in Applied Data Science, University of Michigan. Currently studying genomes and bioinformatics at CUHK.
- 🔭 Right now: shipping mrna-ai-toolkit (codon optimization, neoantigen screening, TrialGPT-style trial matching, scGPT embeddings, LNP chemistry, manufacturability), iterating on deepcatch (cfDNA fragmentomics + methylation → ultra-early multi-cancer detection), and grinding a Lean 4 formalization of Navier–Stokes regularity.
- 🧬 Bioinformatics & computational biology (current focus): mRNA design pipeline (codon, UTR, secondary structure, manufacturability), cancer immunotherapy & neoantigen prediction, cfDNA fragmentomics and methylation signatures, lipid-nanoparticle formulation, scRNA-seq foundation models, AI-driven clinical trial matching — coursework + research at CUHK.
- 📊 Data science & ML: multi-modal medical imaging, ultra-sensitive ctDNA detection (0.001% VAF), fragmentomics feature engineering (FSD, 4-mer motifs, DELFI/WPS, GC-bias correction), NLP for HR analytics (sentiment / emotion / attrition risk), quantitative finance signal design, time-series forecasting, classical ML + deep learning + reinforcement learning — anchored in my M.S. at Michigan.
- 📐 Other research interests: fluid dynamics (Navier–Stokes regularity, Tao averaged, Ladyzhenskaya), formal verification in Lean 4 + Mathlib4, additive combinatorics (Erdős #125).
- 🛠️ Stack: Python, JAX / PyTorch, Lean 4 + Mathlib4, LaTeX (Tectonic), and way too much shell.
- 🌏 Time zone: HKT (UTC+08) — usually awake when half the world is asleep.
- ⚡ Fun fact: My bigger repos are licensed AGPL-3.0-or-later — copyleft, community-first.
Projects I've been actively shipping to:
| Repo | Highlights |
|---|---|
| py-idp | AI document processing (12+ LLM backends, OCR, HITL, FastAPI). Recently: test coverage 87%→92%, 1000-example fuzzers, public idp.console, releases v0.3.5–v0.3.8 |
| nl2pbip | Natural-language → Power BI report generator. Recently: agentic self-reflection loop, Fabric Git-integration metadata, OPC-compliant .pbit archives, object-level security, calculation groups, AI-driven schema advisor |
| deepcatch-methylation | cfDNA CpG methylation cancer detection — most active repo right now |
| employee-voice-analytics | NLP for HR feedback (sentiment / theme / attrition risk) |
| speaksql | NL-to-SQL transpiler — Postgres, MySQL, Snowflake, BigQuery, SQL Server, Databricks, DuckDB, SQLite, SAP HANA |
| deeplethe/utopia (external) | Schema columns carry PK/FK across MySQL/MariaDB · ROW_CAP semantics · blocked-derivations follow the recording axis |
| debpalash/VoiceStudio (external) | Test coverage extension for argos lang_code normalizer (aliases + invalid inputs) |
Bioinformatics & computational biology
| Project | What it is | Stack |
|---|---|---|
| mrna-ai-toolkit | End-to-end mRNA design × AI: codon optimization, neoantigen screening, TrialGPT-style clinical-trial matching, scGPT embeddings, LNP composition, manufacturability checks. AGPL-3.0-or-later. | Python · PyTorch · HF |
| deepcatch | Performance-weighted multi-modal fusion for ultra-early cfDNA cancer detection — fragmentomics, methylation, sequence. | Python · ML |
Formal methods
| Project | What it is | Stack |
|---|---|---|
| ns-blowup | Lean 4 + Mathlib4 formalization toward Navier–Stokes regularity (Ladyzhenskaya, Tao averaged, discrete bridges). 100+ theorems, 0 sorries. | Lean 4 |
| erdos-125 | Formal progress on Erdős #125 (sumset / density). log 3 / log 4 irrationality proved. AGPL-3.0-or-later. | Lean 4 |
Data science, ML & quantitative
| Project | What it is | Stack |
|---|---|---|
| cancer-screening | Ultra-sensitive Stage I cancer screening — ML / DL / RL for 0.001% VAF ctDNA detection. | Python · ML |
| cfdna-fragmentomics-pipeline | Tumor-naive cfDNA pipeline: FSD, 4-mer motifs, DELFI/WPS, GC-bias correction, Cancer-vs-Healthy classifier on real FinaleDB. | Python |
| deepcatch-methylation | CpG methylation-based cfDNA cancer detection (Phase 0: FinaleMe smoke validation). | Python |
| CMC | MICCAI 2024 — Robust Semi-Supervised Multimodal Medical Image Segmentation via Cross Modality Collaboration. | Python |
| employee-voice-analytics | NLP toolkit for HR feedback: sentiment, category, emotion, theme, attrition-risk. Local CLI + Databricks. | Python |
| vmaa · AI-Trader · mr-strategy · vcp-signals | Quantitative trading pipelines: scan / price / risk / execution, VCP pattern detection, mean-reversion, agent-native automation. | Python |
| PH125.9x heart-disease · PH125.9x MovieLens | HarvardX data-science coursework (R). | R |
Applied AI / production systems
| Project | What it is | Stack |
|---|---|---|
| py-idp | AI Intelligent Document Processing: 12+ LLM backends, Pydantic schemas, self-hosted OCR (Nanonets-OCR2-3B), auto-schema discovery, HITL review, prod FastAPI. AGPL-3.0-or-later. | Python · FastAPI |
| scrapex | AI-friendly web scraping for Python — URL + schema in, clean markdown + JSON out. | Python |
| nl2pbip | Natural-language → Power BI report generator. | Python |
| speaksql | NL-to-SQL transpiler — write once, query 9 dialects (Postgres, MySQL, Snowflake, BigQuery, SQL Server, Databricks, DuckDB, SQLite, SAP HANA). | Python |
| MDST-Tutorial | Tutorial materials I authored for Michigan Data Science Team (MDST). | Jupyter |
- 🧪 Collaborative research on cfDNA / mRNA / fragmentomics
- 🛠️ Production LLM / AI tooling partnerships (especially document AI, OCR, applied NLP)
- 📖 Formal methods / Lean 4 collaboration
- 🤝 Mentoring / TA opportunities at the intersection of data science and biology
Drop me a line via GitHub Issues or ORCID.
Real-time stats: github.com/rollroyces?tab=repositories
| 🌟 Pinned repos | 4 flagship projects (see above) |
| 🦈 Achievements | Pull Shark · Quickdraw |
| 🎓 Education | M.S. Applied Data Science (U-Michigan) · genomes & bioinformatics (CUHK, current) |
| 🌍 Region | Hong Kong (HKT, UTC+08) |
| 📦 Public repos | browse all |
Open to interesting collaborations on formal math, mRNA tooling, ultra-early cancer detection, and applied LLM systems.
Last updated · September 2026 ·
rollroyces/README.md