AI Engineering Lab
Yogesh Kuchimanchi · Four small tools for inspecting models, retrieval, evaluation, and inference costs.
Use short, non-sensitive examples. Public calculators do not call a paid model. Results describe the selected input or recorded experiment.
Plan an inference workload
Choose a baseline, adjust inputs, then compare listed monthly costs.
Illustrative baselines; adjust to measured traffic and model needs.
Managed API Workload
Traffic and token inputs drive the API estimates below.
Self-Hosted Compute Setup
Allocated hours drive rental cost; model size drives memory fit.
VRAM estimate = (weights + KV-cache allowance) × 1.20 (decimal GB). The 20% buffer is a planning allowance, not measured activation memory. Set KV-cache for your context length and concurrency; these are not inferred from the API token inputs.
API estimates use 30 days of traffic. Always-on compute uses an approximate 730-hour month; prototype uses 160 hours. Custom retains the current hours for you to edit. Prices alone do not establish capacity for the same traffic. Compute billing assumes one continuous allocation; repeated starts may cost more.
Official API and hardware feeds refresh through a one-hour cache. If a feed fails, the result retains its recorded date and a warning.
GPU memory fit is a planning estimate. It does not establish serving throughput.
Owner-only NVIDIA assistant: disabled. The public calculators are available.
Deep links: ?project=planner, ?project=classifier, ?project=rag, and ?project=evaluation on this app's direct URL.