Project summary

I am Murat Akdemir, an independent self-taught developer based in Türkiye. I am seeking $1,000 for a controlled inference experiment comparing sequential and concurrent coordination among open-weight reasoning, coding and visual computer-interaction agents on authorized defensive cybersecurity sandbox tasks.

Project: https://sinister-ai.xyz

Public documentation: https://github.com/HollyMollyq1/Sinister-ai

Goals and method

The intended public benefit is a reproducible evaluation harness, sanitized tasks, deployment instructions and aggregate results that other open-model builders can inspect and reuse. I will validate model serving and handoffs, establish a sequential baseline, and compare concurrent execution on the same controlled tasks. No activity against live third-party targets is included.

The requirement is exactly four concurrent NVIDIA A100 SXM 80GB GPUs, or four NVIDIA H100 80GB GPUs, with CUDA/Linux and SSH or equivalent direct execution access for my own open-weight models. One node is preferred, not mandatory. API credits do not meet the requirement. The window is 7–14 days at 2–4 active hours daily: 56–224 GPU-hours in total. Temporary storage needs are approximately 0.5–1 TB. Exact model/form factor and four-GPU availability must be verified before booking; no capacity is reserved.

How I will know it worked

I will record task completion rates, handoff failures, latency, peak GPU memory and resource use. The deliverables are sanitized tasks, evaluation code, deployment instructions, aggregate results and a technical demonstration, including failures and limitations. Public components are intended to use an appropriate open-source license; this is a planned scoped contribution, not a claim that the whole project is currently open source. A finding that concurrency does not help will also be reported.

Use of funds

Minimum and maximum request: $1,000. GPU runtime including setup/model loading: up to $896; temporary storage: $50; necessary transfer, transaction costs and contingency: $54. No salary is requested. These are planning allowances, not a provider quotation. The runtime estimate uses a planning ceiling of $4 per GPU-hour for up to 224 GPU-hours. Actual costs and suitable four-GPU availability will be checked before spending. No paid compute will start before funding. If the full experiment cannot fit the approved budget, I will disclose a revised scope before spending.

The work begins after funding, onboarding and compatible capacity are confirmed: days 1–3 for serving/memory checks and baseline validation; days 4–10 for evaluation; days 11–14, if available, for results and documentation.

Team and track record

I am the sole applicant, with no registered company or university/research affiliation. This is an early-stage personal project. There is no completed end-to-end demo, concurrent-model benchmark or released software implementation. The public repository currently contains documentation. GPU access is a bottleneck to implementing and testing this proof of concept.

Failure risks

Main risks include grant/payment eligibility in Türkiye, acceptance as charitable research, suitable four-GPU availability, memory and integration problems, loading overhead and the early implementation stage. If funding or compute is unavailable, the experiment will not proceed as described. Unused funds and material scope changes would be handled under the eventual grant agreement; I do not assume permission to redirect them.

Other funding

No funding or GPU award has been confirmed. Compute support requests are pending with other providers. I will disclose overlapping awards and avoid charging the same expense to multiple funders.

AI assistance disclosure

This proposal was prepared with AI assistance using my stated project facts and resource requirements. No existing demo, benchmark or released implementation is asserted. Consider it an early-stage proposal rather than evidence of completed work.