AI engineer · content systems
Most teams either know content or know code. I sit in between: I design the pipeline that turns an idea into finished, publishable video — script, visuals, voice, captions, distribution — and then I run it in production, not in a demo.
What I do
A pipeline that turns a script or a topic into finished vertical video: scene breakdown, AI imagery, animation, voiceover, word-level captions, export. Built around your format, so the output is yours — not generic AI slop.
The unglamorous machinery around content: comment-to-DM funnels, booking and CRM bots, Telegram mini-apps, dashboards, API integrations. Things that quietly run every day and save someone's afternoon.
You already have a content operation and want to know where AI actually helps — and where it just adds cost and risk. I map the workflow, prototype the two steps worth automating, and hand your team something they can maintain.
Selected work
Every project below is real, shipped, and running — or was until the client took it over.
The client needed a captions.ai-class subtitle editor inside their own product: word-level timing, animated highlighting, full styling control, and an export that matched the preview exactly.
Audio extraction in the browser (ffmpeg.wasm), Whisper transcription with word-level timestamps, a full timeline editor (split / merge / retime / edit), a styling panel with social safe zones, and a server-side render pipeline. Plus on-demand translation into 8 languages and a 24-family commercially-clean font set.
Browser preview is CSS; export is libass. Making them agree pixel-for-pixel took a byte-exact snapshot test over the subtitle generator, so a refactor could never silently change output.
Producing a food short by hand takes hours: shot list, imagery, voiceover, assembly. It doesn't scale past a few videos a week.
A pipeline that takes a voiceover script, reconstructs the actual recipe from it with an LLM, builds a shot list and image prompts, generates a master reference photo plus per-scene 9:16 frames, animates them, and renders the voiceover.
The script is narration, not a recipe. Anything the model infers is explicitly flagged inferred and routed to a human check — so the system is fast without quietly inventing cooking instructions.
The standard creator funnel — "comment LINK and I'll DM it to you" — is manual, and manual means it stops working the moment a video takes off.
A webhook service on the Instagram Graph API: keyword rules per post, automatic public reply plus private DM, signature verification, long-lived token rotation on a cron, and a password-protected panel with live channel metrics.
Platform APIs are half the work: account-level webhook subscriptions, App Review, business verification. Proven end-to-end on live data; the remaining gate is Meta's review queue, not the code.
A nail artist ran her calendar over DMs. Double bookings, missed slots, and a no-show rate she was absorbing personally.
A Telegram mini-app: calendar and client list for the owner, QR-code invites for new clients, automatic reminders at three hours ("still coming?") and one hour before, and an end-of-day summary. Reminders run on a 30-second tick over a database queue, not cron — so a rescheduled slot updates instantly.
Scheduling moved out of the inbox entirely; confirmations became a one-tap answer instead of a conversation.
A veterinary clinic network needed phone-number login without SMS costs — and without the 2FA dead-ends that come with unofficial Telegram clients.
One-time-code delivery through the official Telegram Gateway API. No MTProto session, no user account to get locked out of, no per-message SMS bill for users who already have Telegram.
Chat-based coaching falls apart because the model forgets. A useful coach has to remember what you said three weeks ago.
A Telegram coach with a persistent per-user memory model in private blob storage, so context survives across sessions and the conversation compounds instead of resetting.
How I work
30 minutes. What you make, where the bottleneck actually is, and whether AI is honestly the right tool for it. If it isn't, I'll say so.
A written plan: what gets built, what it costs, what it will and won't do. Fixed price where the scope is clear, hourly where it genuinely isn't.
Weekly working builds, not status reports. You see the real thing running on real data early enough to change direction cheaply.
Deployed, documented, keys in your accounts. I either hand you the keys or keep it running on a retainer — your call, not a lock-in.
About
Content people don't write code. Developers don't make content that lands. I do both, which is why the systems I build survive contact with a real posting schedule: they account for the messy parts — formats that shift, models that hallucinate, platform APIs that change their rules mid-quarter.
I run my own channels as well as clients' work, so I'm testing every pipeline on my own output before I hand it to anyone. That's also the honest reason my estimates are usually right: I've paid for the mistakes already.
Registered sole proprietor, invoicing available. Working with clients across Europe and the US.
Contact
One paragraph is enough: what you make, what's slow, and roughly when you need it. I answer everything within a day.
Write me