← Back to portfolio
Thread 05 / Loose ends

Lab

Things I needed to try for real.

01 / Phygital learning cards

Infinite Atlas

Two experiments in making a physical card continue on a phone.

Made with Codex, GPT Image 2, Kling and Seedance 2.
Jonas / Character studies

Look dev

Early character study of Jonas the mouse
Trickster character study of Jonas the mouse
Wardrobe variations for Jonas the mouse
Daredevil character study of Jonas the mouse
02 / Industrial

Industrial

Built worlds, tested for movement, light and material.

Unreal previs / 5s

Clean staged layers, kept as the motion reference.

Seedance / Material pass / 5s

The same sequence, given ground, roads, light and dust.

Wide aerial starting view of a coastal data center campus

A camera-path test: move from this wide campus view to one exact roof, without losing the route.

Veo / First + last frame / 4s

Closer final frame. The middle turns into a morph.

Seedance / First + end frame / 4s

Cleaner approach. The roof gets rebuilt near the end.

Source aerial frame of the Mactung mountain scene

A material test: keep the mountain and route fixed while daylight and surface treatment change.

Mactung conservative material and daylight variation
Imagegen / Conservative pass
Mactung premium material and daylight variation
Imagegen / Natural pass
03 / Adaptive dubbing

Interview FR Studio

The same 53-second interview: the Russian original and a French adaptation fitted to its timing.

Made with Codex and Gradium for French voice generation.
Original interview / Russian / 53s

Original speech and timing.

French adaptation / A2+ B1- / 53s

Adapted French, two voices and subtitles fitted to the source.

04 / Conversational commerce

Printshop Bot

Telegram assistant for ordering prints on clothing. It can talk freely, but it cannot make up a price or send an unfinished order to a manager.

TypeScript · Mastra · Gemini 3.1 Flash Lite via OpenRouter · Telegram Bot API

Order flow / 6 steps One message travels from Telegram to a confirmed order.

Architecture diagram: a Telegram message enters through a webhook, retains session state, moves through a router and slot parser, checks catalogue and pricing rules, then becomes a confirmed lead and a Telegram reply.

Two routes / one checkpoint Clear replies stay in code. Ambiguous requests go through AI.

Decision diagram: clear replies use the fast Slot Filler path, while ambiguous requests go to the AI Order Agent. Both converge on SKU, catalogue and confirmation checks before a price or lead is created.
Try it in Telegram ↗