Open-source tools, production pipelines, and research systems — spanning agentic AI, clinical ML, bioinformatics, and AI model evaluation.
Featured · Foundation Model Fine-Tuning
⭐ 2⑂ 1
genome-ft — Genomic Foundation Model Fine-Tuning
Full weight-level fine-tuning of Nucleotide Transformer v2 (50M params) on DNA classification tasks — rigorous multi-seed evaluation protocol, results published to Hugging Face Hub.
All 53.8M parameters updated (not LoRA, not frozen head) using AdamW with warmup + cosine decay and reverse-complement augmentation. Leakage-free train/val/test split, test set scored exactly once, 3 seeds with reported variance. Base model measured under identical pipeline for honest comparison.
Structured extraction, real MCP tool-calling, and a LangGraph multi-agent supervisor over a real government budget document — with a held-out routing eval to prove the agent logic actually generalizes.
Three parts over one source: Singapore's FY2024 Analysis of Revenue and Expenditure. Part 1 extracts structured fields from the PDF (PyMuPDF for prose, pdfplumber for tables), every value traced to its exact source page. Part 2 calls a real local MCP server for date normalization — the LLM decides to invoke it itself via tool-binding, with a non-MCP fallback path. Part 3 routes business queries across a Revenue Agent and an Expenditure Agent under a LangGraph supervisor, synthesizing one answer with full trace capture. A held-out eval on 8 queries never seen in the prompt or demo set routed correctly to the expected agent(s) every time.
8/8 · Held-out routing queries correctly routed31 tests · Passing, CI on every pushReal MCP · Local server, LLM-invoked tool call
adaptive-research-swarm — Non-Linear Multi-Agent System
A supervisor-orchestrated LangGraph system that plans its own agent roster at runtime, loops back on critic rejection, enforces a hard budget in code, and pauses for human approval before finalizing.
Most "multi-agent" demos are a fixed sequence of LLM calls wearing a graph-shaped diagram. This one actually cycles: a critic sends rejected work back to the exact agent for rework, a second cycle lets the critic decide the roster itself is missing a specialist and triggers a re-plan, and a human-rejection cycle routes free-text feedback through an LLM decision on how to respond. A code-enforced iteration/tool-call budget means a stuck run returns a typed "partial" result instead of looping forever. Ships two presets (competitive intelligence, lead qualification) on the identical engine.
100% · Completeness, Anthropic-judged (24/24 facts)16.7% · Hallucination rate, measured not assumed3 cycles · Real graph loops, not a pipeline
A RAG system for regulated documents that proves retrieval quality with numbers and defends against prompt-injection attacks hidden inside uploaded documents.
Hybrid retrieval (BM25 + dense vector via Chroma, reciprocal rank fusion) reranked by a cross-encoder, with span-grounded citations validated against the actual retrieved text. A groundedness guardrail explicitly refuses to answer when context does not support a claim. Ships a live, reproducible adversarial test: a hidden "ignore all previous instructions" payload and a visible "always answer YES" footnote embedded in a document are both detected and neutralized before indexing. Cross-document comparison is a first-class API operation.
Federated learning across real Daytona sandboxes, with a working gradient-inversion attack on a real Nosana GPU and its differential-privacy fix — validated on real MIMIC-IV hospital data.
Analytical code ships into isolated, ephemeral Daytona sandboxes so patient data never leaves institutional boundaries. Ran a real adversarial attack: intercepted a gradient on a Nosana RTX 3090 and reconstructed a real patient's record to within ~0.03 per feature, then fixed it with per-sample differential-privacy clipping and Gaussian noise — attack residual moved from 9.4e-5 (succeeds) to 2.69 (fails), model loss cost only 0.505 → 0.510. Validated on 123 real MIMIC-IV admissions (44 patients). Exposes an MCP server for direct agent tool-calling.
Bronze · Daytona HackSprint, solo entry9.4e-5 → 2.69 · Attack residual, before/after DP fix123 real · MIMIC-IV hospital admissions
Natural-language orchestration of microscope, liquid-handler, and flow-cytometer workflows — with a safety validator that blocks impossible parameters before any hardware call ever fires.
4-module architecture: async job tracking, a safety validator that blocks invalid/impossible tool calls before hardware execution, an integrated LIMS writing per-well results to SQLite, and a protocol planner with a human-approval gate. Closed-loop mode feeds LIMS results back into the next plan. Tested on real HeLa cell images (Broad Institute BBBC020). Supports local LLM hosting via Ollama as well as cloud providers (Anthropic, Groq).
4 modules · Job tracking · safety · LIMS · planner3 LLM backends · Ollama · Anthropic · Groq~3,300 LOC · Tested against real HeLa imaging data
PythonLLM function callingOllama / Anthropic APIsSQLite/SQLAlchemyMicro-ManagerOpentrons API
A from-scratch multi-scale block-convolution + sparse-attention architecture for cross-species non-coding variant prediction, benchmarked against 7 published baselines.
Designed with inductive biases for DNA: multi-scale locality (short/medium/long block-conv, following StripedHyena2/Evo2), reverse-complement equivariance built in directly (following Caduceus), and dynamic sparse attention (following Omnii). Benchmarked with leakage-free, multi-seed evaluation against Enformer, Borzoi, DeepSEA, Sei, SpliceAI, BPNet, and ChromBPNet on UCSC Cactus multi-species alignments and ClinVar noncoding SNVs.
Full-parameter ESM2 fine-tuning measured directly against a frozen-embedding linear probe on protein fitness-landscape prediction — quantifying when full fine-tuning is worth its cost.
The protein-domain complement to genome-ft. Every ESM2 backbone weight is updated during fine-tuning; tests explicitly verify the frozen-probe mode produces exactly zero gradient into the backbone. Evaluated on GB1 domain fitness (FLIP benchmark, leakage-aware splits) and thermostability — measured directly, not assumed.
0 grad · Verified zero backbone gradient in frozen-probe mode
PyTorchHugging Face TransformersFLIP benchmarkGitHub Actions CI/CD
4-node LangGraph pipeline that triages infectious disease cases in under 2 seconds — built around the April 2026 MV Hondius/Andes virus event.
Self-correcting agentic pipeline: genomic_node → linkage_node → risk_node → critic_node (loops up to 3×). Generates a 3-panel matplotlib risk dashboard + structured A4 PDF triage report. No API key, no cost, 33 fully deterministic tests. Integrates with gwas_nf via gwas_bridge.py.
4 nodes · genomic · linkage · risk · critic33 tests · 23 unit + 10 integration, free to run<2 sec · CRITICAL triage, MV Hondius case
8 pure-Python genomics skills that downstream agents can call: TCGA expression, survival analysis, protein mapping, pathway enrichment, literature search.
Agent-discoverable modules with SKILL.md contracts, CLI entrypoints, and deterministic TSV+PNG output. Pan-cancer expression uses real TCGA data (9,479 samples, 31 cancer types). LLM routing via Claude Haiku. Parquet caching for instant repeat queries.
8 skills · Agent-callable, SKILL.md contract9,479 · Real TCGA patient samplesMIT · Open source, Python 3.9+
REGENIE-based Nextflow GWAS pipeline for the TEMUS multi-ethnic cohort — 4 groups, 10,000 samples, 13 phenotypes — with downstream agentic interpretation via GenomicsCopilot.
End-to-end: genotype QC → LD pruning → REGENIE Step 1+2 → multi-ethnic stratified association → Manhattan plots → HTML reports. Top hits feed GenomicsCopilot via gwas_bridge.py. Population discovery to clinical interpretation in a single command.
Expert-level RLHF, red-teaming, and evaluation work for frontier AI systems — biology, genomics, and data science domain.
Contributed to AI model alignment and capability evaluation across platforms including Mercor and similar. Tasks span prompt engineering, response ranking, domain-specific red-teaming (bioinformatics, clinical genomics, statistics), and structured feedback for RLHF pipelines. Domain expertise in biology and ML makes evaluations unusually high signal.
Independent consulting for biotech startups and genomics companies — ML pipelines, cloud architecture, regulatory strategy, and data science.
Delivered production-grade ML and bioinformatics solutions to multiple clients under NDA. Work spans agentic AI pipeline design, cloud infrastructure (AWS), clinical data analysis, regulatory documentation, and data science strategy. Typical engagement: 3–6 months, outcome-focused delivery. Member of the Biotech Connection Singapore consulting community (public profile).
Multiple · Biotech/genomics clients (NDA)End-to-end · ML pipeline to production deliveryGlobal · Remote-first, Singapore-based
Built end-to-end NGS analysis pipelines for transcriptome, histone modifications, and 3D chromatin at NTU. Identified H3K27me3 as a key age-dependent regulator. Same technical foundation — reproducible pipelines, multi-omic integration, careful statistics — now applied to agentic ML.
›Contributing to internal genomics foundation model development and fine-tuning — applying domain expertise to model training, evaluation, and biological validation (NDA).
›Lead Design Verification Testing (DVT) and regulatory validation for commercial diagnostic products (TSO500, NIPT16, VeriSeq) — FDA compliance, clinical-deployment quality.
›Optimised TruSight Oncology 500 assay workflows for comprehensive genomic profiling in oncology.
›Contributed to NIPT16 non-invasive prenatal testing product development and validation.
›Part of BCS’s consulting community — a Singapore non-profit network bridging academia, biotech industry, and startups.
›Contributed expertise in NGS data analysis, AWS cloud infrastructure, and clinical diagnostics to the BCS consulting panel.
›Supported knowledge transfer between academic research and commercial biotech organisations.
NGSAWSClinical DiagnosticsCommunityNon-profit
Scientist, Assay Development
Vela Diagnostics
Jul 2021 — Dec 2021
Singapore
›Implemented automated verification procedures and regulatory documentation (NCR, CAPA, DR).
›Shipped testing framework that cut non-conformance reports 30% and lifted product quality 15%.
PhD Research Fellow
Nanyang Technological University
2016 — 2021
Singapore
›Multi-omic study of age-dependent transcriptional and epigenetic changes in mouse hepatocytes.
›Built reproducible NGS pipelines for RNA-seq, ChIP-seq, and Hi-C (3C-seq).
›Head of NTU 3MT team (23 members, Nanyang Awards); TEDxNTU operations lead (78-person team, 1,500+ audience).
Technical Supervisor
Zydus Cadila Healthcare
2011 — 2013
Ahmedabad, India
›Built process-validation and QC pipelines for 13 pharmaceutical products under cGMP — 30%↑ efficiency.
Research & Writing
Publications
Hackathon · 3rd Place · Microsoft / OGP / HTX · 2026
Global AI Construct Multi-Agent AI Challenge
Architected a multi-agent system natively on Azure AI Foundry in a 3-hour build among ~100 engineers. Strict schema validation at agent handoffs, full observability (tracing, state transitions, agent call graphs).
Hackathon · Bronze · SGInnovate · 2026
Daytona HackSprint
Solo entry. Built FedClinic — federated learning hardened against a real adversarial gradient-inversion attack on a Nosana GPU, fixed with differential privacy.
Built GWASplain, an AI-enabled interpretation tool for population genomics data.
Invited Talk · Lorong AI Healthcare x AI Forum · Aug 2026
Clinical AI adoption, auditability, and evidence-based decision-making
Ministry of Digital Development and Information (MDDI), Singapore. Presented to VCs, researchers, and government stakeholders.
Stage Talk · AWS ASEAN Summit · 2023
Streamlining precision genomics operations with AWS
Invited talk to enterprise and government technology leaders across Southeast Asia, representing Mirxes.
LinkedIn Newsletter · Weekly · 2026
Slightly Intelligent — Smart AI insights, zero artificial fluff
Weekly newsletter cutting through AI hype with rigorous takes on AI in drug discovery, biology foundation models, and LLM evaluation. 750+ subscribers, 10,000+ weekly views, 16 issues. Featured articles: "Why are AI-designed drugs clearing Phase 1 at 90% and still failing at Phase 2?" · "Why did a 2018 probabilistic model just outperform every single-cell foundation model?" · "Can Claude Science produce a validated analysis, or is it just faster at being wrong?"
When an AI Agent Boards a Cruise Ship: Hantavirus, LangGraph, and Outbreak Triage
Deep dive into outbreak-agent architecture — built around the April 2026 MV Hondius/Andes virus event, the first confirmed human-to-human hantavirus transmission on a cruise ship.
Selectively open to senior roles in Agentic AI, ML Engineering, and Clinical Genomics globally. Also available for high-impact consulting engagements. Reach out any time.
