You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
I want someone to describe a place, step inside it, and learn with a guide who understands what they are looking at. They should be able to change the world by speaking. That is the ambition behind Sciescape.
A convincing scene can still have an unusable doorway or objects that change when revisited. The guide needs to see the same world as the visitor. I am asking for $25,000 to build and test an open component that keeps these parts consistent, with $5,000 for a smaller pilot.
I am Navish Kumar, an employed doctoral researcher at Basel. Sciescape is my personal project, developed with coding agents, without university sponsorship, an advisor, faculty PI or institutional signatory.
Five panoramas and their provenance are public. The checkpoint/resume utility passed 23 CPU tests on 15 November 2026. Geometry completed only 2 of 21 development cases, which informed the work and are not independent validation. The full walking, lifelike avatar and natural voice experience is unfinished. External adoption is not established.
Rooms, openings and objects will have stable identities in one versioned scene, shared by rendering, movement, edits and the guide. Edits will be checked before display; failed edits will leave the previous state intact. A public viewer, synthetic scenes and replay tools will work independently of the private application; saved-scene replay will need no paid API.
Over 12 weeks I will reproduce failures, pilot baselines, build the component and run a fixed comparison. After pilots, I will freeze 21 new scene requests, two seeds, scripts, scoring and software versions. Four conditions give 168 planned journeys: the current depth-panorama pipeline, shared state, separate interaction state, and a runnable, compatibly licensed public baseline. Any scope reduction precedes this freeze.
Tests will check passable routes, object identity on return, persistent edits and guide claims against independent renderer evidence. Results will include visual quality, latency, memory, cost and uncertainty by scene. Timeouts and refusals count as failures. An independent engineer, not yet appointed, will reproduce the runner and review scoring.
Success means a tool others can run and better complete journeys, with quality and cost trade-offs reported. Text2Room, DreamScene360 and explicit scene representations already exist. I will test whether this combination helps and publish negative results too.
$16,200: my research and implementation, 180 hours at $90.
$4,800: independent engineering and reproduction, 60 hours at $80.
$3,500: GPU and host compute.
$500: storage and transfer.
This proposes 15 hours weekly alongside my PhD. The figures are allowances, not quotes. Spending starts only after funding and terms are agreed. Checkpoints, CPU replay, bounded retries and automatic shutdown will keep work within agreed funds and terms.
The $5,000 pilot would cover 40 of my hours ($3,600), 10 hours of independent review ($800), compute ($500), and storage/transfer ($100). It would produce a small prototype and failure replay across six scenes, two conditions and one seed: 12 exploratory journeys. It would not fund the full 168-journey study, a new avatar or the complete voice experience. Intermediate funding would require a revised scope and budget before work starts.
Shared state may preserve faulty geometry or cost too much speed and visual quality. Compute, model behaviour and licences may limit comparisons. I will test these risks early and preserve useful replay tools even if the hypothesis fails. Synthetic scenes and licensed voices avoid private recordings and real-person likenesses. Learning gains and human acceptance are outside this study.
New funded code and documentation will be Apache-2.0. Examples and evaluation materials will use suitable open permissions, subject to upstream rights. The tools will run independently of private Sciescape code. Credentials, restricted assets and third-party model weights are excluded. The film keeps its separate noncommercial media licence.
Prototyping has used promotional Google Cloud credits and Kaggle/Colab allocations; I have not reconciled their monetary value. Applications to Lambda, CloudRift, fal, OpenAI, EGI, Yotta, Inference.net and Sequoia OSS are pending. Sentient has the same $25,000 work package for preliminary review. I will disclose decisions and remove overlapping costs or agree separate work packages before accepting funds. No cost will be charged twice.
AI helped research this opportunity and draft and edit this proposal from my records and instructions. It does not validate the research or establish the proposed results.