A community for engineers shipping production AI.
Ask hard engineering questions. Share what you actually built — not a polished demo. Discuss RAG, agents, MCP, evaluation, architecture and production AI with people who do this for a living.
Ask an engineering question
Stuck on a RAG quality issue? Agent won't stop looping? Tool calling returning junk? Bring the symptom, your evidence and what you've already tried.
Ask a questionShare what you built
Built a RAG system, an agent, an MCP server, an eval harness? Share the architecture, the trade-offs you made, and what you would do differently.
Share your buildFeatured prompts
Cost blowup? Infinite loops? Tool-call errors? Hallucinated function arguments? Share the symptom, what you found, and the guardrail you added afterwards.
Not the conference-talk version — the real one with the cache, the reranker, the eval pipeline and the human-review queue. What did you cut from the idealised version, and why?
Honest takes only. What would have to be true for MCP to matter in 12 months? What would kill it?
Recent discussions
Categories
Retrieval, chunking, embeddings, hybrid search, reranking, faithfulness, citations.
Agent loops, tool calling, planning, memory, multi-agent systems, guardrails.
Model Context Protocol — building tool servers, clients and integrations.
Golden sets, LLM-as-judge, faithfulness, regression gates, online eval.
System design, trade-offs, failure modes, scaling, cost modelling.
Latency, cost, reliability, observability, security, deployment.
Forward Deployed Engineering — customer-facing AI engineering in the field.
Roles, interviews, transitioning into AI engineering, salary, growth.
Share what you built, get feedback, find collaborators, show your work.
- Describe the symptom — what is happening, when, and how often.
- Share your evidence — traces, metrics, logs, eval scores. Screenshots of dashboards are great.
- Say what you've already tried. "Restarted the service" counts. So does "read the docs and the source code".
- Name the versions — model, library, framework. "It doesn't work" is not enough.
- When someone answers, mark the answer. It helps the next engineer who hits the same problem.
Ground rules
- • Be an engineer first. Egos and credentials don't debug your system.
- • No vendor pitches disguised as questions or answers.
- • Disagree with the idea, not the person.
- • Cite sources. Link to the docs, the paper or the code.
- • When you solve your own question, post the answer anyway — someone else has the same problem.
Want community in your inbox?
AI Engineering Weekly includes a community question of the week — one real engineering problem to think about.
We don't fake discussion counts or upvotes. If a category is empty, it's empty. The community is small and real.