Hi, it's Owen Loh

PourDynamics engine

A multiscale physics engine mapping brewing variables to coffee flavour.

Two coupled physics: Darcy flow of water through the porous coffee bed (advection–dispersion), and a 2D-axisymmetric Doyle–Fuller–Newman extraction model ported from lithium-ion batteries — coffee and a battery are, mathematically, the same equations — validated to sub-0.2%. A sensory layer then maps concentrations to taste. Drag the two controls and read off J = dy/dx: which knobs actually move the cup.

• Darcy + DFN
• simulation
• flagship

Catalon (startup cofounder)

AI automation for chemical commercial teams. Purchase-order processing, ~40 min → ~3 sec.

startup cofounder · founded, built & exited

As startup cofounder I built the agentic pipeline behind Catalon. It reads an inbound PO email, classifies intent, resolves the customer against their ERP, extracts and price-validates every line item — chemical-synonym resolution, density-based lbs↔gal conversion, customer rules — and drafts an ERP-ready order for one-click approval. Sub-3-second latency, on top of existing systems, no rip-and-replace.

• Multi-stage state machine: deterministic prefilter → classify → SQL+LLM customer-resolve → schema extraction → self-healing retry — with circuit breaker, rate limiter and dead-letter queue.
• Schema-decoupled: a reusable core with hot-swappable per-customer “context cartridges.” Eval-driven tuning, not fine-tuning → 99%+ accuracy.
• 2 paying distributors, $60K ARR and a $1.2M pipeline within 2 months — then exited.

• multi-agent
• founder
• exited

SeisPilot

Voice + natural language → 3D seismic navigation inside pro software “Tornado.”

Professional seismic software has a brutal learning curve and expert time is scarce. I wrapped it in an agent: describe what you want to see and an LLM turns it into live, validated commands against the application’s API — like a seismologist driving it for you.

• NL → JSON-RPC via an LLM, gated by a command validator (only whitelisted calls run), with multi-LLM fallback and a queued listener.
• 20-level undo/redo; bookmark engine; bidirectional transform between seismic (inline/crossline/depth) and Cartesian space.
• Built solo in ~3 weeks at Viridien on my own initiative; drew immediate traction from teams across China, the Netherlands, the US and Singapore.

• MCP
• seismic
• agentic

iGEM 2025 (Oxford University)

Gold medal — high-performance CRISPR guide discovery.

Oxford University team · Grand Jamboree, Paris 2025

The job: find guide RNAs that hit a target bacterium while sparing the commensal flora — a nearest-neighbour search by Hamming distance over millions of genomic sites. I built a two-layer engine: a trie prefix filter (exact first 13 nt) drops ~85% of candidates, then a BK-tree Hamming search over a 2-bit-packed genome handles the rest.

• Brute force is O(N·M·L) — every target guide against every commensal guide. The hybrid is O(N·(k + log M·L)): linear in targets, with the database term dropping from M to log M.
• ~2.3M indexed guides across 50 genomes; ~3 s per guide; ~2000× faster than the BLAST equivalent; 1.2 GB bit-packed vs 4.8 GB as strings; 100% accuracy. ML & algorithm lead at iGEM 2025.

• BK-tree + trie
• bit-packed DNA
• CUDA

Battery State-of-Health

Imperial College London · Faraday Institution · ongoing publication

A cell’s health is usually read by tearing it down. I treat it as an inverse problem instead: reconstruct the full-cell OCV from its cathode and anode contributions, and the way they have shifted reveals exactly which electrode lost capacity — non-invasively. The standard fit misbehaves for NMC811–graphite because scipy.curve_fit returns a local minimum.

• Differential Evolution for a true global fit — least-squares 0.0004 vs 0.005.
• DV_fit: differential-voltage feature-matching via optimised Dynamic Time Warping — peak/dip locations are invariant in lithiation.
• Degraded_OCV_fit over a Pchip-DTW manifold — the most consistent modes so far. NMC811–graphite, 41 cells; awaiting experimental verification.