Selected work

PROBLEM Turning an article into a watchable video was a manual editing job measured in hours, which meant publishers simply did not do it.

APPROACH A Go pipeline that parses the page, scripts it with an LLM behind a three-provider routing layer, assembles scenes from stock, generated and templated assets, then renders through After Effects and FFmpeg across a fleet of AWS workers fed by SQS — spread wide because a single render node was nowhere near fast enough — with a Vue editor over the top for anyone who wants the last word.

PROBLEM Air cargo paperwork moved on paper between airlines, warehouses and the federal customs service — every handoff a place to lose a day.

APPROACH A PHP/PostgreSQL platform speaking IATA CIMP, SITA, ARINC and EDIFly Type B, with GOST-qualified digital signatures through CryptoPro and direct integration into the customs IT system — built with a team of four, then extended to a second country's customs regime.

PROBLEM Ad and content experiments needed backends that stayed available under spiky publisher traffic without degrading response times.

APPROACH Go and PHP microservices with Vue front ends, built around measurable request budgets and backpressure rather than optimistic caching, provisioned in Terraform and watched in Grafana.

PROBLEM Campaign managers were repricing hundreds of ads by hand across dozens of sites according to shifting inventory margins.

APPROACH A rules engine over the Yandex.Direct API that applied a manager's pricing intent automatically and published one advert to many sites.

PROBLEM My wife runs an Instagram account about our dog and wanted to put text and stickers onto photos she already had. Every editor simple enough that she would actually use it wanted a monthly subscription for that handful of features.

APPROACH A browser image editor built on a real layer model — pixel, text and adjustment layers, blend modes, clipping masks — running client-side on a canvas engine so the editing never leaves the machine, with only the AI work (background removal, generative fill, upscaling) pushed out to serverless GPU. Split across two origins: a static marketing site on the apex domain so it can be indexed, and the editor SPA on a subdomain where none of that matters.

PROBLEM The small jobs around video — change a container, grab a still, get a clip under Discord's size limit, work out how many frames are in ten seconds at 23.976 — are mostly served by sites that want you to upload the file to someone else's server first and wait.

APPROACH A browser toolkit that uploads nothing, picking one of three engines per job rather than reaching for the heaviest: copy the streams untouched where the codec already plays in the target container, WebCodecs where re-encoding is genuinely unavoidable, and ffmpeg.wasm only as the fallback — downloading a whole media engine to rewrite a container header is work nobody should pay for. The size-targeted compressor derives its bitrate from the clip's own duration and audio track, then measures the finished file and re-encodes when it overshoots.