Royce
I'm an ML × bioinformatics engineer working at the intersection of biology and AI. I build end-to-end systems for mRNA design, cfDNA cancer detection, and production AI tooling — LLM backends, OCR, NL→SQL, document AI.
Open to roles
What I'm looking for
- ML Engineer — applied AI / LLM systems
- Bioinformatics / Computational Biology Engineer — mRNA design, liquid biopsy, foundation models for biology
- Research Engineer — ML-for-biology, scRNA-seq, foundation models
- Part-time / consulting on cfDNA, mRNA tooling, formal methods
Featured work
deepcatch
Multi-modal cfDNA cancer detection combining fragmentomics, methylation, and sequence features.
View project →mrnavax
End-to-end mRNA design: codon optimization, UTR engineering, neoantigen prediction, LNP formulation.
View project →py-idp
Production document AI · 23 LLM backends · HITL review · 92% test coverage.
View project →speaksql
NL→SQL transpiler across 9 SQL dialects with semantic validation.
View project →nl2pbip
Natural language → Power BI report generator from a single prompt.
View project →ns-blowup
Lean 4 + Mathlib4 formalization toward Navier–Stokes regularity.
View project →What I work with
Selected work
Bioinformatics & computational biology
- cfdna-fragmentomics-pipeline—Tumor-naive cfDNA features (FSD, 4-mer motifs, DELFI/WPS, GC-correction).
- cancer-screening—ML / DL / RL for 0.001% VAF ctDNA detection.
- CMC—MICCAI 2024 · robust semi-supervised multimodal medical image segmentation.
Applied AI & production systems
- scrapex—AI-friendly web scraper (URL + schema → clean markdown/JSON).
- employee-voice-analytics—HR NLP toolkit (sentiment, emotion, theme, attrition risk).
Quantitative
- vcp-signals · AI-Trader · vmaa — VCP patterns, mean-reversion, agent-native execution.
Writing
Long-form writing on ML for biology, applied AI tooling, and formal methods. Every draft and the full source live in rollroyces/blog — 7 posts, ~9,300 words.
- Multi-modal cfDNA fragmentomics for ultra-early cancer detection—Performance-weighted fusion of fragmentomics, methylation, and sequence.
- 9 SQL dialects from one grammar—How speakSQL transpiles natural language to vendor-specific SQL.
- 100 theorems toward Navier–Stokes regularity in Lean 4—Formalizing fluid dynamics with Lean 4 and Mathlib4.
- Production document AI—23 LLM backends, HITL review, and per-field eval reports.
- Agentic Power BI generation—From one natural-language prompt to a Git-friendly .pbip workspace.
- Eleven tools and twelve model adapters—A usable Python layer over the mRNA-therapy AI literature.
- Turning exit interviews into attrition signals—A dual-runtime HR-NLP toolkit with privacy guardrails.