Linkedin tech templates

How to reduce API response times in a production app

Slow API responses are almost always caused by the same five things. Here is where to look, in order. 1. N+1 queries. If your API makes a database call inside a loop, you are likely making hundreds of queries where one would do. Add query logging and look for repeated patterns. 2. Missing indexes. Check your query execution plans. A missing index on a frequently-filtered column can turn a fast query into a slow one at scale. 3. Synchronous work that should be async. Sending emails, generating thumbnails, and calling third-party APIs inside a request handler all add latency that should happen in a background job. 4. No caching layer. Identical requests for the same data should not hit the database twice in the same minute. 5. Payload size. Strip fields the client does not need. Serialization time adds up faster than most people expect. Investigate in this order. The first two fix most production slowdowns without touching the architecture.

The energy cost of AI model training is growing rapidly

Training a single large language model can consume more electricity than 100 US households use in an entire year. That number was startling two years ago. Now it describes models that are already considered mid-tier. The newest frontier models require significantly more, and we are training more of them, more often, for more use cases. This is not an argument against AI development. It is an argument for taking the infrastructure question seriously much earlier than most companies currently do. The teams building AI applications in 2025 that are not thinking about energy, cooling, and compute efficiency are building on an assumption, that the cost structure will stay manageable, that may not hold. The technical debt in AI right now is increasingly physical, not just in the code.

post: why most companies are still not getting value from AI tools they've bought

Most AI tool adoption fails the same way: the software gets purchased, a few people use it inconsistently, and six months later it's a line item no one can justify. The problem is rarely the tool. It's the workflow. AI amplifies whatever process it's plugged into. A broken or undefined process doesn't become efficient with AI, it becomes a faster broken process. The companies seeing real returns aren't the ones who bought the best tools. They're the ones who mapped their workflows before buying anything.