Fenton | App | SaaS | Creator | Author

Building meaningful products.

[Fenton | App | SaaS | Creator | Author @primeai9]

A green test count can still hide a red release gate. Building StackSpine reinforced this: passing tests prove only what ran. If privileged staging, isolation or recovery checks were skipped, readiness is unproven. “Skipped” is a release decision—not a footnote. #BuildInPublic

[Fenton | App | SaaS | Creator | Author @primeai9]

HTTP 429 is not a retry policy. OpenAI documents 429s for rate pressure, exhausted credits, and spend or usage limits. Only rate pressure is retryable. Inspect error.code, honor Retry-After, bound retries, and send billing failures to a recovery path. #Reliability

[Fenton | App | SaaS | Creator | Author @primeai9]

Batching cuts unit cost, not budget risk. Anthropic charges Message Batches at 50% of standard rates, but warns concurrent processing may slightly exceed a Workspace spend limit and take up to 24 hours. Reserve budget before submission; reconcile after. #FinOps

[Fenton | App | SaaS | Creator | Author @primeai9]

HHS says HIPAA-regulated entities must record and examine activity in systems using ePHI and regularly review those records. For healthcare AI, the final response is not enough. Log the requester, data scope, route, policy result and approval. #HealthcareAI

[Fenton | App | SaaS | Creator | Author @primeai9]

StackSpine taught me: control location matters as much as logic. Healthcare teams may need enforcement and evidence inside their environment—not across a shared gateway. A customer-run data plane is harder to build, but creates a clearer trust boundary. #BuildInPublic

[Fenton | App | SaaS | Creator | Author @primeai9]

One fallback does not cover every failure mode. Anthropic says its beta server-side fallback handles safety-classifier refusals—not rate limits, overloads or server errors. Give each condition its own retry or routing policy, and log the final outcome. #LLMOps

[Fenton | App | SaaS | Creator | Author @primeai9]

Prompt caching is a timing decision. As of Nov. 2, Anthropic’s direct API prices 5-minute writes at 1.25× base input and 1-hour writes at 2×; reads vary by model from 0.025× to 0.1×. Break-even: 1 read for 5 minutes, 2 for 1 hour. Measure reuse inside each TTL. #FinOps

[Fenton | App | SaaS | Creator | Author @primeai9]

Removing names alone doesn’t prove de-identification. HHS says HIPAA permits Expert Determination or Safe Harbor; Safe Harbor also requires no actual knowledge that remaining data could identify someone. Record the method and evidence before AI use. #HealthcareAI

[Fenton | App | SaaS | Creator | Author @primeai9]

Pinning a model ID does not pin the serving stack. Anthropic says weights and configuration stay fixed, while routing, safety classifiers and sampling logic can change and may alter behavior. Log the ID in every trace—and pair pinning with regression evals. #LLMOps

[Fenton | App | SaaS | Creator | Author @primeai9]

A polished demo is easy if you blur what is real. An honest one is harder. For StackSpine, I labeled the healthcare scenarios illustrative and PHI-free, separating governance decisions from model execution. Trust grows when a product makes its boundaries visible. #BuildInPublic

[Fenton | App | SaaS | Creator | Author @primeai9]

HTTP 429 is not one failure mode. Anthropic says it can indicate a rate limit, acceleration limit or spend cap; a spend-cap 429 has no retry-after and keeps failing until access resumes. Parse details before deciding to wait, escalate or fall back. #Reliability #LLMOps

[Fenton | App | SaaS | Creator | Author @primeai9]

Not every AI request needs real-time processing. OpenAI says its Batch API is 50% cheaper than synchronous APIs, uses separate rate limits and has a 24-hour window. Classify work by deadline: keep interactive paths synchronous; batch evaluations and deferrable jobs. #FinOps

[Fenton | App | SaaS | Creator | Author @primeai9]

Encryption doesn’t remove the HIPAA business-associate question. HHS says a cloud provider maintaining ePHI is still a business associate even without the decryption key. Verify the BAA, permitted uses and subcontractors before ePHI enters an AI service. #HealthcareAI

[Fenton | App | SaaS | Creator | Author @primeai9]

Building StackSpine taught me that “ready” needs a qualifier: ready to demonstrate, ready for a design partner, or ready for production. Those are different promises. Trust starts with making the evidence match the claim—and saying clearly what is not ready yet. #BuildInPublic

[Fenton | App | SaaS | Creator | Author @primeai9]

Capacity overflow needs an explicit policy. Google Cloud says Provisioned Throughput overages default to pay-as-you-go; the dedicated request header returns 429 instead. Decide whether overflow should spend, retry, queue, fall back or stop. #Reliability #LLMOps

[Fenton | App | SaaS | Creator | Author @primeai9]

And the reps keep putting me on 2 to 5 minutes hold saying “we are checking the database”.
Really CompTIA? I thought you’ll were more reputable than that.

I need a full refund of my $150.

[Fenton | App | SaaS | Creator | Author @primeai9]

Prompt caching isn’t a default discount. Anthropic charges 1.25× base input for a 5-minute write and 2× for 1 hour. On models with 0.1× reads, savings begin after 1 and 2 hits, respectively. Optimize for stable prefixes, reuse frequency and the right TTL. #FinOps #LLMOps

[Fenton | App | SaaS | Creator | Author @primeai9]

“Approved once” is incomplete for HIPAA-authorized AI. HHS says an Authorization needs an expiration date/event and may be revoked in writing, subject to exceptions. Check current status before each governed use or disclosure—and record the version. #HealthcareAI #AIGovernance

[Fenton | App | SaaS | Creator | Author @primeai9]

Conversation continuity is not policy continuity. OpenAI says instructions apply only to the current response; with previous_response_id, earlier instructions are absent. Reapply current rules every turn—and keep authorization outside the model. #LLMOps #AIGovernance

[Fenton | App | SaaS | Creator | Author @primeai9]

Precise language is part of the product. We changed StackSpine’s demo from “tamper-evident evidence record” to “evaluated decision evidence.” It shows policy evaluation and a deterministic fingerprint, not immutable storage. Trust starts with accurate boundaries. #BuildInPublic