Fidelity under hard constraints.

Pirated live streams need to be traced back to the account that leaked them.

The mark must survive re-encoding, cropping and screen capture while staying invisible to viewers.

Led the R&D of a watermarking algorithm combining signal processing, computer vision and ML decoding, and built most of the low-latency streaming backend around it.

Industry-leading decode robustness. Infrastructure cost cut by about 40% and compute by about 90% while usage tripled.

A major sports broadcaster needed AI built into its production workflow.

Four months to delivery, on infrastructure that had to be auditable and finely permissioned.

Sole architect of 450+ Terraform-managed AWS resources, a relationship-based permission system and a GPU-bound ML image pipeline.

Delivered on time and now part of the client's daily workflow.

Enterprise clients need access rules finer than roles: per entity, per scope, and across accounts.

Roles, groups, inheritance and single sign-on all have to compose, and permission checks still have to be fast.

An IAM-like, relationship-based model on OpenFGA and DynamoDB with policy-based roles, scope inheritance, group membership and cross-account sharing.

One permission layer behind every part of the platform, including SSO.

Engineering standards written for people do not reach the AI coding agents that now write much of the code.

Every rule has to be specific enough for an agent to apply and for a check to verify.

Rebuilt 33 design and standards documents as a rule system: about 100 rule IDs across 10+ domains, each with a severity, the file paths it covers and how it is enforced (lint, CI or AI review).

One set of standards that people and agents both read, apply and are checked against.

Operators needed to ask plain-language questions of a large audit history.

Answers in seconds, at a cost per query that stays negligible as the log grows.

Retrieval-augmented generation with prompt size held constant regardless of history length.

Sub-10-second answers at a fraction of a cent per query.

Teaching material needs to be annotated, organised and searched directly on a canvas.

One person building every layer, from schema and infrastructure to interface, pricing and go-to-market.

A custom camera system, SVG annotation primitives with keyboard-driven placement, semantic search with synonym expansion, and tiered content gating.

In development, with infrastructure cost modelled across three scaling scenarios.

Writing fingerings into a score is slow, and for music like Opus Clavicembalisticum it takes days. There is no dataset that pairs notes with the fingerings pianists actually use.

No labelled data. Each fingering depends on the fingers just used and the notes still to come, and the two hands mirror each other.

Generated the training data from music theory: every scale mode and arpeggio inversion with comfort-optimised fingerings, extended by weighted random walks over a transition map. A two-layer GRU per hand (PyTorch) with circular pitch encoding feeds its own predictions back as input. Reads and writes MusicXML.

80–90% validation accuracy; fingers a Bach fugue in under half a second.

Finding every subject entry by hand in fugues that run from ten minutes to two hours, as Sorabji's do.

Entries are transposed, partial or altered, and may be reversed, inverted, augmented or diminished.

Fuzzy matching on intervals and durations with a custom edit distance (costs for contracted intervals, inserted or dropped notes, and rests), a sliding window that trims each match from both ends, and weighted interval scheduling to choose the best non-overlapping matches across transformations.

Marks subject entries and five transformations directly in the MusicXML score, in roughly linear time in the length of each voice.