On 2026‑10‑14 I posted about a new two‑part article series that teaches Kubernetes fundamentals for AI engineers, and I announced the upcoming End‑to‑End AI Engineering Bootcamp.
The first article in the series is completely free.
Link to the first article: https://lnkd.in/ed2HfHjV
The next bootcamp cohort starts on 2026-10-21, a week after the posting date.
AI Observability is a must have in your tool belt as an AI Engineer.
Additional metadata like input token count is persisted with the span so that we can estimate the cost of the procedure.
Cost for calling LLM APIs will be variable depending on the length of inputs and produced outputs. You would usually trace this information and analyze it to help forecast expenses.
GenAI systems are non-deterministic and will deteriorate over time. They need to be evaluated on span level rather than input/output of the entire system so that you can tune each piece separately.
I decided to adopt tracing as a core practice for all my RAG pipelines, instrumenting each major step as a span and persisting token‑count metadata to estimate costs and spot performance regressions.
MCP Client – Clients that maintain 1:1 connections with servers.
MCP Server – Lightweight programs that each expose specific capabilities through the standardised Model Context Protocol.
Local Data Sources – Your computer’s files, databases, and services that MCP servers can securely access.
Remote Data Sources – External systems available over the internet (e.g., through APIs) that MCP servers can connect to.
- Agents (MCP Hosts) would implement and communicate via A2A protocol, that enables:
These components together let an MCP Host act as an autonomous agent, filling the gap where MCP alone does not handle multi‑agent state sharing.
- Secure Collaboration.
- Task and State Management.
- User Experience Negotiation.
- Capability discovery - similar to MCP tools.
- Retrieval chunking option: small vs large chunks (slide/tumble windows optional).
- Retrieval window type: sliding or tumbling.
- Embedding model: contextual embeddings for documents and queries.
- Vector DB hosting: choose managed service or self‑hosted deployment.
- Metadata stored with embeddings: timestamps, source, tags.
- Similarity measure: cosine similarity (common) or ANN‑first vs metadata‑first query path.
- Heuristics applied: time importance, reranking, diversity ranking, source retrieval preferences.
- LLM selection: proprietary API vs self‑hosted open‑source (performance converged).
- Prompt engineering: required to steer output, avoid jailbreaks, and align with retrieved context.
- Observability: continuous monitoring, evaluation, and security scanning in production.
I also advertised the upcoming End‑to‑End AI Engineering Bootcamp, which starts on October 19 2026, offering a 15 % discount with code YEAREND15.
I treat the ML Training pipeline as a first‑class artifact in CI/CD.
I consider the CI/CD pipeline a separate entity from the Machine Learning Training pipeline (true).
In production the pipeline is ready for Continuous Training (enabled).
The upcoming End‑to‑End AI Engineering Bootcamp starts on October 19 2026 with a 15 % discount using code YEAREND15.
On October 5 2026 I posted a LinkedIn article for AI engineers building agentic systems in enterprises. I summarized Anthropic’s top workflow patterns for agentic systems, explained when each is valuable, and added practical deployment tips.
- Prompt Chaining – decompose a complex task into sequential LLM calls; higher accuracy but added latency.
Use code YEAREND15 for 15% off – valid for the next two weeks.
Parallelization: Initial input is split into multiple queries to be passed to the LLM, then the answers are aggregated to produce the final answer.
Prompt Chaining would be combined with following patterns, a pattern replace an LLM Call node in Prompt Chaining pattern.
I recommend trying simpler workflow patterns first; in heavy‑production pipelines Prompt Chaining is often paired with other patterns (e.g., routing after a chain step) to balance accuracy and latency. ✨
- Routing – classify input and dispatch down predefined paths; enables specialized sub‑workflows.
- Parallelization – split input into independent queries, run concurrently, then aggregate results; improves throughput and can boost accuracy via voting.
- Orchestrator – a dedicated LLM dynamically breaks a task into sub‑tasks and delegates to other LLMs or sub‑workflows; handles non‑static topologies.
- Evaluator‑Optimizer – one LLM generates a result, a second LLM evaluates and feeds back for refinement; supports iterative improvement.