AI + SYSTEMS / FIELD NOTES
Practical patterns for human-led agents, model serving, and infrastructure choices.
01 / AGENTS
Example patterns for delegating bounded work while keeping people accountable for decisions and changes.
Share goals, constraints, and existing design
Ask for options, risks, and open questions
Verify assumptions and choose a direction
Capture the decision and rationale
02 / INFERENCE
Local inference and AWS Inferentia have different cost and operations profiles. The right fit depends on the model and traffic.
| Decision factor | Local inference | AWS Inferentia |
|---|---|---|
| Good fit | Prototyping, private data, edge devices, or existing hardware | AWS-hosted inference with a compatible model and sustained demand |
| Cost shape | Hardware, power, and maintenance; low marginal cost when already provisioned | Instance and data-transfer charges; utilization strongly affects unit cost |
| Operations | Own drivers, runtime, capacity, and hardware availability | Own AWS instance, deployment, scaling, and Neuron runtime compatibility |
| Model path | Choose a runtime supported by the target CPU, GPU, or accelerator | Validate framework, operators, and export path with AWS Neuron |
Qualitative comparison, not a benchmark. Validate model compatibility and measure your own workload before choosing.
03 / COMPUTER VISION
Explore the stages from input frame to human review. The annotations are illustrative; no model runs in the browser.
Illustrated sample frame · static annotations, not model output
YOLO / FRAME
A sample street scene enters the pipeline. In a real service, validate format, dimensions, and source before processing.
04 / ARCHITECTURE
Start with what the workload needs. Compare cost shape and operational effort before picking a hosting model.
Serve prebuilt pages from a CDN and avoid running an application server for requests that do not need one.
| Trade-off | StaticSuggested | Server | Hybrid |
|---|---|---|---|
| Cost shape | Storage, requests, and data transfer; usually little idle compute. | Compute and managed services; baseline capacity can cost while idle. | CDN and storage plus API or compute usage for dynamic paths. |
| Operations | Low: build and publish files; no app server to patch or scale. | Higher: runtime updates, capacity, security, and availability. | Moderate: static delivery plus a smaller server-side surface. |
| Typical fit | Portfolios, documentation, and content with build-time data. | APIs, authenticated sessions, and server-side workloads. | Content-first products with selected dynamic features. |
Cost depends on region, traffic, storage, and data transfer. Treat this as a design prompt, not a cloud quote.