Blog | Hokstad Consulting

Blog

AI Agent Tool Recovery: Retry vs Re-route vs Replan

Choose the smallest safe recovery for failed tool calls: retry, re-route or replan using error type, execution state, risk and budget.

Read more

How RAG Systems Resolve Conflicting Source Evidence

How RAG systems detect and reconcile conflicting sources by comparing claims, grouping evidence by scope and authority, and verifying citations.

Read more

TAPAS: Thermal- and Power-Aware Scheduling for LLMs

Details TAPAS's placement, routing and VM reconfiguration to reduce GPU thermal and power peaks, plus pilot and telemetry checks.

Read more

Kubernetes Scale-to-Zero: Cold Starts vs Idle GPU Cost

Scale-to-zero saves GPU cost only when nodes terminate; weigh idle-cost savings against cold-start latency, model load time and node-retention.

Read more

Tool Contract Tests vs End-to-End Tests for AI Agents

Catch interface drift with contract tests, then use sandboxed end-to-end tests to verify agent decisions, state and safe recovery.

Read more

Generative AI Output End-User Licence Agreements

Grant only the rights your product can actually deliver; map output, input and liability controls before UK AI launch.

Read more

Pydantic Validates AI Tool Arguments Before Execution

Validate AI-generated tool arguments with Pydantic v2 using strict types, safe correction feedback and two retry attempts before execution.

Read more

Memory Retrieval Tests Track Agent Answer Quality

A correct answer doesn't prove memory retrieval — test retrieval, answers, latency and cost separately before release.

Read more

AI Agent Recovery: Replay and Duplicate Tool Actions

Save action IDs, reuse idempotency keys, limit retries and reconcile unknown outcomes to prevent duplicate payments.

Read more

BackdoorBench: Attack and Defence Testing for AI Teams

Benchmark image-classification backdoors with repeatable tests measuring ASR, clean accuracy and compute cost across attacks and defences.

Read more

Kubernetes LLM Networking: 7 RDMA Deployment Checks

Seven RDMA checks to verify device registration, pod access, GPU–NIC/NUMA locality, runtime access, connectivity and transfer performance.

Read more

Speculative Tool Execution: When Does It Pay Off?

Speculative tool calls save latency only when slow, predictable reads outweigh wasted work—pilot read‑only tests and measure accuracy and cost.

Read more