Project summary

I am Murat Akdemir, an independent developer in Türkiye. I am requesting $1,000 to run a small, controlled evaluation of Sinister AI, an early-stage assistant for authorized cybersecurity research. The study needs four concurrent NVIDIA GPUs with at least 80 GB VRAM per GPU. It is inference work, not model training.

Goals and method

I want to compare sequential and concurrent execution on a fixed set of defensive sandbox tasks. I will establish a baseline, repeat the same tasks with concurrent inference, and record completion, latency, handoff failures and resource use. All tests will be confined to isolated systems where I have explicit authorization. No activity against unauthorized third-party targets is proposed.

How I will know it worked

A useful outcome is a completed, measured evaluation, whether or not concurrency improves performance. I will report aggregate results, limitations and failure cases in a high-level public summary. I am not claiming that a completed end-to-end benchmark already exists. This request does not promise release of proprietary source code, private model identities, internal architecture or private datasets.

Compute and budget

The requirement is four A100 80 GB or H100 80 GB GPUs concurrently, CUDA/Linux with SSH or equivalent direct container execution, and about 1 TB of temporary storage. A single GPU or hosted-model API credits cannot run the planned evaluation. I plan 90–140 aggregate GPU-hours across four GPUs: about 22.5–35 active wall-clock hours, within a 7–14 day window after funding and a compatible allocation are confirmed. Setup and model-loading time must fit within that allowance.

Requested funding: minimum $1,000; maximum $1,000. Planning ceilings are $840 for GPU runtime, $80 for temporary storage and $80 for necessary transfer and transaction costs. These are budget allowances, not a provider quotation. I will verify current pricing, availability and total cost before booking. No salary or hardware purchase is requested. If the study cannot fit the approved budget, I will disclose the limitation and agree any material scope change before spending. No paid compute will be launched before funding is available.

Team and current status

I am the sole applicant, a self-taught developer without a registered company or university affiliation. The project is at an early prototype and evaluation stage. This request is for the compute needed to carry out and measure the proposed study; it is not a claim of production readiness or validated performance.

Public benefit and privacy

The intended contribution is an honest high-level account of whether concurrent AI assistance improves reliability and efficiency in controlled defensive research. Positive and negative observations can help other independent builders understand practical limitations. Private implementation materials are outside the proposed public outputs.

Risks and failure

The main risks are integration errors, insufficient memory, unstable execution, no measurable benefit, and unavailable four-GPU capacity. If the experiment fails, I will report what could and could not be measured rather than claim a successful demonstration. Unused funds and any material changes would be handled under the eventual grant agreement.

Other funding

No GPU funding award has been confirmed in the records supporting this request. Compute-support applications have been sent to other providers. I will disclose overlapping awards and will not charge the same expense to more than one funder.

AI assistance disclosure

This proposal was prepared with AI assistance from my stated project requirements. It does not assert completed benchmarks, publications or an existing compute allocation.