Writing & Research

Notes from the build.

Three places I write, plus working papers from the research side of the practice. Pick a channel — they each lean a different way.

Channels

Substack

Long-form essays on building data platforms, distributed systems, and the engineering behind founder ventures.

Systems thinkingFounder notesArchitecture

Medium

Practitioner write-ups — patterns from production: entity resolution, event-driven backends, ML in the loop.

BackendData engineeringML

Hashnode

Developer-first deep dives — code-heavy posts on TypeScript, Python, infra, and the tools I ship with.

TypeScriptPythonDevOps

Research

Working papers & technical notes.

Zenodo (v1)Apr 2026

GenMol: A Lightweight Variational Autoencoder for Target-Aware Drug-Like Molecule Generation with Conditional Property Control

A compact SMILES-based VAE for target-aware molecular generation with conditional control over drug-likeness properties (QED, logP, MW). Designed as the generative module powering NeoRx.

VAEMolecule generationDrug discovery

Read paper

Zenodo (v1)Mar 2026

Disease-Type-Aware Biological Intelligence Layers for Preventing Symptom Marker False Positives in Computational Drug Target Identification

Biological intelligence layers that distinguish causal disease drivers from symptom markers, eliminating a major class of false positives in computational target identification pipelines.

Drug discoveryCausal inferenceBiomedical

Read paper

Zenodo (v1)Mar 2026

CausalBioRL: Reinforcement Learning with Causal World Models for Autonomous Drug Discovery

An RL framework that learns causal world models over biomedical environments, enabling agents to reason about interventions rather than correlations during autonomous drug discovery.

Reinforcement learningCausal inferenceDrug discovery

Read paper

Zenodo (v1) · Open source on GitHubMar 2026

NeoRx: Causal Inference for Automated Drug Target Identification via Pearl's Do-Calculus over Multi-Source Biomedical Knowledge Graphs

End-to-end causal pipeline applying Pearl's do-calculus across 8 biomedical knowledge graphs. SMILES BiGRU VAE (4.1M params, 97% validity), hierarchical RL (UCB1 + CEM), AutoDock Vina docking. Mean F₁ = 0.474 across 7 disease validations with zero false positives.

Causal inferenceKnowledge graphsDrug discovery

Read paper