Insights & Updates
Explore insights from our engineering team on software development, AI, cloud architecture, and industry best practices.
Master Claude API integration for enterprise applications. Learn authentication, streaming responses, tool use, vision capabilities, and best practices for building reliable AI-powered features with Anthropic's most advanced models.
Jishu Labs
CTO
January 20, 2026
13 min read
Featured Article
23 articles tagged “llm” · page 1 of 2
Clear tag filterPrompt injection is when untrusted text reaching a model's context gets treated as instruction rather than data. It has held OWASP's number one LLM risk slot across every edition, and agents made it materially worse.
AI features have a cost structure traditional software does not: it scales with usage, varies per request, and is set by design decisions made months earlier. A method for estimating before you build.
Serving a 7B model is roughly 10-30x cheaper than a frontier model for tasks where accuracy is equivalent. The engineering question is which tasks those are, and how to find out without guessing.
Chat history is not memory. A memory layer is durable, retrievable state about decisions, preferences and facts that survives past the context window. What belongs in one, and what should stay in a log.
Long context windows and better tool use took work away from retrieval-augmented generation. RAG did not become obsolete — its job got narrower. A decision framework for when to retrieve, when to load, and when to call a tool.
Prompt wording stopped being the bottleneck around 2025. The harder problem — deciding what lands in the context window at every turn — got a name. What context engineering is, and the four things it covers.
Traditional observability tells you an agent responded. It does not tell you whether the response was correct. What to instrument for agentic systems, and how tracing and evaluation fit together.
Asking a model politely for JSON produces JSON most of the time, and most of the time is not a contract. How constrained decoding, schemas and validation combine into something you can build on.
Benchmarks say your agent works; production says otherwise. A practical evaluation approach covering end-to-end, trajectory and component scoring, and how to build an eval suite from your own failures.
Multi-agent architectures are the default recommendation of 2026 and the wrong answer for most teams. A decision framework for choosing between one agent, a planner-executor split, and a full orchestration graph.
Guardrails are the controls that sit around a model rather than inside it. Where each type belongs, what each can and cannot catch, and why buying beats building here.
Three ways to make a model behave the way you need, routinely treated as competitors when they solve different problems. What each actually changes, and the order to try them in.
Get the latest insights on software development, AI, and technology trends delivered to your inbox.
Get consistent results. Collaborate in real-time.
Build Intelligent Apps. Work with Jishu Labs.