Filter by topic (137)
  • Fine-tune vs RAG vs prompt (how to choose)

    A decision guide for choosing prompting, RAG or fine-tuning based on whether the gap is knowledge, behavior or an under-specified task.

    Fine-tuning · AI EngineeringUpdated 2 min read

  • Fine-tuning (full vs LoRA / PEFT)

    What full fine-tuning, LoRA and QLoRA each update, what they cost in memory and storage, and when adapters over a shared base model are the better choice.

    Fine-tuning · AI EngineeringUpdated 2 min read

  • Function / Tool Calling

    How an LLM calls external code through a validated JSON payload, why runtime schema checks matter, and how to debug a tool-call failure rate.

    Agents & orchestration · AI EngineeringUpdated 2 min read

  • Golden datasets & regression testing

    How to build a version-controlled golden dataset, run it in CI to block regressions in prompts, models and RAG, and grow it from production incidents.

    Evaluation & quality · AI EngineeringUpdated 2 min read

  • GraphRAG / Knowledge Graphs for Retrieval

    How GraphRAG builds an entity graph from a corpus, retrieves connected subgraphs for multi-document questions, and where extraction cost and errors bite.

    Retrieval & RAG · AI EngineeringUpdated 2 min read

  • Guardrails & Policy Engines (Input/Output)

    How input and output guardrails such as NeMo Guardrails and Llama Guard check prompts and responses, why prompt rules fail, and how to tune false alarms.

    Safety & governance · AI EngineeringUpdated 2 min read

  • Hallucination detection & grounding

    How grounding ties answers to retrieved documents, and how a claim-by-claim faithfulness check catches unsupported statements.

    Evaluation & quality · AI EngineeringUpdated 3 min read

  • Human-in-the-Loop & Approval Gates

    How approval gates pause an agent before irreversible actions, which tools to gate, and what happens when the gate itself misfires.

    Agents & orchestration · AI EngineeringUpdated 3 min read

  • Hybrid Search (Keyword + Vector)

    How hybrid search runs BM25 and vector indexes together, merges the ranked lists, and why each index alone misses queries.

    Retrieval & RAG · AI EngineeringUpdated 2 min read

  • Idempotency & Retries with LLM Calls

    How idempotency keys, backoff and upserts make it safe to retry an LLM call without duplicate charges, emails or writes.

    Reliability · AI EngineeringUpdated 2 min read

  • Inference serving engines (vLLM, TGI)

    How vLLM and TGI use continuous batching and a paged KV cache to serve self-hosted models efficiently, and when you need one.

    Serving & infrastructure · AI EngineeringUpdated 2 min read

  • Instruction tuning

    How supervised fine-tuning on instruction-response pairs turns a base model into an instruct model, and when light SFT is worth doing.

    Fine-tuning · AI EngineeringUpdated 3 min read

  • Labeling & annotation

    How rubrics, annotator agreement and adjudication produce reliable labels, and where model-assisted labeling is safe to use.

    Data & pipelines · AI EngineeringUpdated 2 min read

  • Latency, throughput & batching (TTFT vs tokens/sec)

    How prefill and decode set time to first token and tokens per second, and why batch size pulls the two in opposite directions.

    Serving & infrastructure · AI EngineeringUpdated 2 min read

  • LLM gateways / proxies (e.g. LiteLLM)

    How an LLM gateway such as LiteLLM unifies provider APIs, centralizes keys, budgets and PII redaction, and where it can become a single point of failure.

    Serving & infrastructure · AI EngineeringUpdated 2 min read

  • LLM-as-judge

    How an LLM scores outputs against a rubric, which biases to correct, and when a deterministic grader or a human is the better judge.

    Evaluation & quality · AI EngineeringUpdated 2 min read

  • LLMOps vs MLOps vs DevOps

    DevOps, MLOps and LLMOps differ in what they ship and how they test it. This article covers prompts, eval sets, token cost and prompt injection risk.

    LLMOps · AI EngineeringUpdated 2 min read

  • Long-Context vs RAG

    Compare stuffing a corpus into a 100k to 1M token window with retrieving a small slice through RAG, including cost, latency and lost-in-the-middle recall.

    Retrieval & RAG · AI EngineeringUpdated 2 min read

  • MCP (Model Context Protocol)

    How the Model Context Protocol lets one tool server work with any MCP client, and why it cuts N times M integrations to N plus M.

    Agents & orchestration · AI EngineeringUpdated 2 min read

  • Mixture-of-Experts (MoE)

    How mixture-of-experts models route each token to a few experts, why VRAM still holds every expert, and how router load imbalance shows up in production.

    Emerging · AI EngineeringUpdated 2 min read

  • Model & prompt versioning & rollback

    Pin model snapshots, tag prompt and RAG index versions as one bundle, log the bundle per request, and roll back by repointing to the prior bundle.

    LLMOps · AI EngineeringUpdated 2 min read

  • Model routing, fallback & load balancing

    How routing, fallback and load balancing choose a model, survive 429 and 5xx errors, spread traffic across keys, and avoid a single hardcoded model.

    Serving & infrastructure · AI EngineeringUpdated 2 min read

  • Multi-Agent Systems & A2A

    How supervisor and specialist agents split work, what Agent-to-Agent protocols standardize, and when coordination overhead favors a single agent.

    Agents & orchestration · AI EngineeringUpdated 2 min read

  • Multimodality (vision/audio)

    How multimodal models like GPT-4o, Claude and Gemini encode images and audio as tokens, reason across them, and when Whisper or OCR still fit better.

    Model foundations · AI EngineeringUpdated 2 min read