Tian Pan
Software Engineer
The Fallback Model Whose System Prompt Was Tuned for Someone Else
Failover keeps your LLM app available when the primary goes down — but the fallback model reads a system prompt that was tuned for someone else, and your users notice the difference.
The Few-Shot Example Your Model Treated as Binding Precedent
Few-shot examples are not neutral demonstrations — they are case law. Models bind to the closest example by surface tokens and inherit its constraints, shipping confident-wrong answers eval suites cannot see.
The Fine-Tune Artifact Your Departing Engineer Took With Them
A fine-tuned model is not a file in a registry; it is the closure of a pipeline over a training set. The teams that ship only the weights discover their bus factor the day a base-model migration arrives and the original engineer is gone.
The JSON Schema Your Output Passed and Your Downstream Consumer Rejected for Semantic Drift
A JSON schema validates shape, not meaning. When an LLM upgrade shifts the distribution of values inside a still-valid schema, downstream consumers detonate while the producer's dashboard stays green.
The KV Cache Eviction Your Provider Called Cache Pressure and Your Bill Called a Doubled Prefix Charge
Prompt caching looks like a configured discount, but KV cache eviction on shared LLM infrastructure turns it into a probabilistic one — the same conversation can cost several times more on a busy day with no code change.
The KV Cache Warm-Up Cron That Ran in Blue and Never in Green Because the Host Pinning Never Moved
A blue/green deployment orphaned a cron pinned to the old color, the prompt cache went cold, and the bill quietly tripled — anatomy of a silent regression and the four practices that close the seam.
The Legal Disclaimer That Leaked From The Answer Into The Tool Call Arguments
A safety disclaimer added to your system prompt does not stop at the user-facing reply. It rides along into every tool call argument the model produces — and into the downstream systems those calls fire against.
The LLM-as-Judge Ensemble That Agreed Because All Judges Were the Same Family
An LLM-as-judge ensemble drawn from one provider family measures family-internal consistency, not judgment quality, and the high agreement number is an artifact of provider selection nobody named.
The Logprobs Field Your Provider Removed That Broke Your Confidence Router Silently
A confidence router that stopped escalating low-confidence answers, the silent provider tier change that caused it, and how response-shape contracts, population-level alerts, and a fallback written for the wrong failure mode hide together.
The max_tokens Default Your Provider Raised That Doubled Your Tail Response Length
An LLM provider quietly raised the default max_tokens value and your p99 output length doubled overnight. The parameter you did not send is the configuration that changed under you — here is how to stop inheriting defaults you do not control.
The MCP Server Your Team Forgot Was Running with Prod Credentials
An MCP server running on a developer laptop with a CI-grade OAuth token is a production attack surface. Here is how DNS rebinding, bad bindings, and shared tokens turn one compromised tab into a deploy-key leak.
The Model Card Benchmark Whose Methodology Shifted While Your Contract Cited the Number
A benchmark number is a measurement under a protocol, and the protocol is what your vendor controls. Pin the methodology, or contract on your own eval suite.