LUKA ADAMIA Watch the trailer →

Filmmaker and engineer in Tbilisi. I make films with the same models Sundream runs on, and I build the agents, renderers and pipelines behind them.

Five brand films · one GPU renderer · 20+ systems in production
Remote from Tbilisi (UTC+4). New York mornings are my afternoons.
See the work → First 90 days

Brief in. Film out.
The decisions stay yours.

That is the shape of everything I build. A structured document is the source of truth, an agent edits it, a renderer turns it into frames, and nothing that costs money runs until a person says yes. Hand it a brief and a budget and it keeps going; step in whenever you like.

Knight · concept trailer· storyboardSpecBoardRender
StoryboardSpecMedia Production Script · 6 scenesCast · 2 elementsFrames · 6 of 6Clips · 4 of 6Edit · 0:32
All scenes6 scenes · 0:32
Scene 1The rider in the tiger skin, river at dawn.
Scene 2Riders crest the snowfield, banners low.
Scene 3Close on the face. Hold two beats.
Scene 4Title over black, low strings.
Scene 5Alternate take for the cut-down.
Scene 6End card with the live link.
Assistant
YouCut a 30-second version for Instagram. Keep the rider consistent.
AgentScenes 1, 2, 3 and 4. Two clips still need renders, both from the rider's references.
AgentRun the two renders?
ApproveChange
Estimate shown before anything runs. Nothing spends without a yes.

Make it. Build it. Ship it.

Three things I do every week, each one a piece of what Sundream does.

01

I make films with these models.

A concept trailer for The Knight in the Panther's Skin, a 2:40 battle film and a toy-brick series, all generated with Seedance. Keeping one character the same across hundreds of shots is the problem I fight every day.

02

I build the engines behind them.

VIVI turns a written spec into a finished video; my WebGPU renderer runs text-motion scenes at 96.6 fps in 4K, checked frame by frame. ASFI is a chat that asks only what it needs, then renders the ad.

03

I ship them to real users.

Twenty-plus systems in production: a tourism CRM with 63 tables, a Georgian legal AI with about twenty paying accounts, storefronts run by an agent with forty tools.

The Knight · concept trailer
L.A99 · live 3D, my own build
XRA · in production
XRA product page

Every piece of Sundream,
already in my hands.

Switch between how I fit and the numbers behind it, without leaving the page.

Fit · proof
Pieces that match
8

Hover a piece to keep its details on screen.

Elements · consistent cast

Lived it: one character held across a 2:40 battle film and a trailer.

Assistant · asks first

A 40-tool agent where every write waits for a yes, can be undone, and has a daily cap.

Bring your own agent

MCP servers I build and run every day.

Many models

Primary and fallback models, key rotation on errors, loops that always end.

Server-side export

My own renderer: spec in, 4K MP4 out, hardware-encoded.

Dialogue and sound

Text-to-speech with word-level timing, so words land on the cut.

Credit estimates

Token budgets that pause a run instead of overspending.

Tested, not assumed

Every claim gated: SSIM on frames, adversarial checks on data.

GPU renderer
96.6fps at 4K
Pipeline speed-up
393→82seconds
Frame match
0.989SSIM, full ad
Systems live
20+in production
One schema
63tables
Admin agent
40tools
Content engine
52workers
Team
15hires, one month

Three things I'd build first.
One way to start.

A paid trial task first, then whatever fits.

Ideas from using tools like Sundream as a filmmaker. Switch to see how each would work, and tell me which matters most.

Idea 01Quality

Consistency check

after every shot
scored per element
WhyA clip that drifts from its character costs credits and trust. Catch it before the next scene builds on it.HowA vision model scores each new clip against the element's reference images and flags drift, scene by scene.
Start with this
  • Scored per character and location
  • Flags before more credits are spent
  • Feeds back to the assistant
Idea 02Speed

Instant previews

before the render
scrub the cut in seconds
WhyPeople decide on the cut by watching it. Waiting on a server render for every tweak slows that down.HowA GPU preview of the timeline in the browser, built on what I learned writing VIVI's renderer. Pay only for the export you keep.
Start with this
  • Same frames as the final export
  • Server render only on export
  • Works on long timelines
Idea 03Sound

Dialogue on the beat

voice to frame
word-level timing
WhyLong-form film lives on dialogue. Lines, captions and cuts should land together without hand-timing.HowText-to-speech with word timings from forced alignment, so captions and cuts snap to the spoken word.
Start with this
  • Word timings on the timeline
  • Captions that follow the voice
  • Cuts that snap to lines
How we workCustom

Let's talk

trial first
contract or full-time
WhyThe fastest way to judge me is real work on your code, with a deadline.HowContract through my own company, so there is no visa or payroll overhead; full-time if it fits.
Send a trial task
  • A paid 1–2 week trial task
  • New York mornings, my afternoons
  • Straight about gaps: about two years shipping professionally

Numbers on this page come from my own projects and are measured, not estimated. Ideas are proposals from the outside; I would check them against your roadmap and data before building anything.

Let's make the
next cut together.

Send me the hardest open problem on your board. I'll come back with a plan and a first version.

Write to me →
Remote · contract or full-time · available in two weeks