Budget Allocation
Where every share of the budget goes.
We are deliberately running lean — no inflated line items, no vendor markup. Most of the budget goes to where it actually matters: GPU compute and dataset access.
Allocation Breakdown
Six categories, one disciplined plan.
Percentages reflect approximate share of total project resources.
% of total project resources
GPU Compute
40%Reserved compute hours for fine-tuning runs and evaluation. Largest single line item — model training is GPU-intensive.
Licensed Data Buffer
23%Reserved capacity to license premium academic datasets in Phase 2 (subject to negotiated TDM agreements).
Contingency
17%Risk buffer for unexpected costs — dataset licensing surprises, additional compute, or scope adjustments.
Software & Tooling
10%Experiment tracking (W&B), evaluation frameworks, and other essential dev tooling.
Cloud Storage
7%Object storage for raw datasets, training checkpoints, and model artifacts during development.
Open Data Acquisition
3%Bulk data transfer fees for arXiv S3 and similar open-access sources — minimal cost.
Lean by Design
How we are stretching every dollar.
Open-weight foundation
Starting from a free open-weight base model means zero licensing fees — vs paying for proprietary APIs.
Parameter-efficient fine-tuning
LoRA / QLoRA cuts training cost by 90%+ vs full fine-tuning — same quality, fraction of the GPU hours.
Free academic datasets first
arXiv, Semantic Scholar, PubMed Central, OpenAlex — high-quality scholarly data at near-zero cost.
In-house development
No agency markup. No outsourced consultants. Built by SCSU faculty and students who understand the problem.
Smart contingency
A meaningful contingency reserve protects against scope shifts and licensing surprises.
Designed to scale
Same architecture serves 1 user or 10,000. Scaling up is an infrastructure decision — not a rebuild.
Lean development. Limitless scalability.
Running a project this lean is a feature, not a constraint. Because HuskyBot is built on open-source foundations — not subscription APIs or proprietary platforms — scaling up means adding hardware, not rewriting agreements. There is no per-seat license that explodes with adoption, no API rate limit that throttles growth, no vendor that can change terms on us.
Build once. Serve everyone. Forever.
See how we are getting funded.
Outreach to GPU providers, dataset vendors, and stakeholders is in motion.
Outreach & Email Drafts