Evals
Product management for AI features: the test set is the spec
AI features don't fit strict test cases. What worked for us: real examples, good and bad answers, and a pass bar agreed with the client before building.
· Prafull B.
Insights
Notes from production.
What we learn building AI systems that have to work: the parts demos skip, and the decisions that matter once real users and real volumes arrive.
AI engineering
Prompt, context, loop engineering: why the names keep changing
Prompt engineering, then context engineering, now loop engineering. Each name came when the last one hit its limit. One skill stayed the same underneath.
· Prafull B.
AI engineering
What changes when an AI agent gets a million requests
At scale the first limit is the model provider's quota, not your servers. Tokens per minute, shared rate limits, fair queues and cost per request.
· Prafull B.
AI engineering
Security for agentic workflows: put the rules in code
Prompts don't enforce permissions. Where the rules for an AI agent live, the three-part test for injection risk, and why every step must be safe to repeat.
· Prafull B.
AI engineering
Why so many AI agents ship as CLI tools
The terminal already has every tool an agent needs: git, tests, files and every CLI you installed. Why agents ship there, and how to keep shell access safe.
· Prafull B.
AI engineering
Server-side vs client-side agents: where should yours run?
Where an AI agent runs decides whose keys it holds, whose data it sees and who pays for the tokens. Backend code vs frontend code, with the same trust rules.
· Prafull B.
AI engineering
Why production agents are rare, and what has to exist first
Most AI agents in demos are workflows, and that is fine. What keeps real agents out of production is the code around the model: permissions, logs, limits, undo.
· Prafull B.
AI engineering
Why AI features need background jobs
One click became 10-15 model calls and requests ran for minutes. What moved to a worker, what broke when it crashed, and the small version to build first.
· Prafull B.
AI engineering
Why your chat feature needs SSE, and when to use WebSockets
A slow AI answer behind a spinner looks broken. How SSE fixes it, when WebSockets fit better, what proxies do to streams, and why long work belongs in a job.
· Prafull B.
AI engineering
What you build around an LLM API call in your backend
One endpoint, one model call is a demo. A production feature needs a schema, domain checks, timeouts, a log of every call and a way to say I don't know.
· Prafull B.
AI engineering
What is AI engineering? Lessons from a chatbot that fell flat
AI engineering starts with an API call. The work is finding where a probabilistic feature fails on real data, and defining requirements that let you check it.
· Prafull B.
Have an AI system to deliver?
Tell us what your client needs, or what your team is stuck on. We reply within two working days.
Talk to us