Field notes on AI automation — agents built with n8n and LangGraph, RAG systems, evals, and LLMOps. No hype, no demos: just architecture decisions, tradeoffs, and lessons learned from real projects.
The archive
Keep withdrawn RAG evidence out after restores and delayed ingestion: independent tombstones, generation checks, cache provenance and a fail-closed release gate.
LFM2.5-24B-A2B brings local reasoning and tool use. Verify the entire agent data boundary, commercial terms and task outcomes before trusting a private deployment.
Audit inspection reference labels before scoring a model: blinded repeat reviews, disagreement matrices, prevalence-sensitive kappa and an adjudicated holdout.
Missing canary traffic is not a health signal. Build explicit pass, fail and inconclusive gates with tested Argo provider and Prometheus fixtures, matched workloads and fresh denominators.
Define attacker access, valid perturbations and eligible denominators before reporting robustness. An executed synthetic classifier test separates new errors, coverage and abstention.
More inference workers can increase waiting rather than throughput. An executed ONNX CPU experiment separates request concurrency, session pools, BLAS limits and spinning.
A real citation can still point to the wrong maintenance revision. Resolve asset scope, effective time and authority before ranking evidence, with synthetic negative tests.
A synthetic two-tone test shows how sensor downsampling invents a low-frequency signal. Define anti-aliasing, causal filter state and feature timing before reducing rates.
Bound inference metric labels before they exhaust the observability budget. A pinned OpenTelemetry fixture measures exported series and reconciles overflow without hiding lost detail.
Reconcile test reports, assessed purpose and monitoring plans with the actual AI release. A synthetic validator separates structural consistency from legal sufficiency.
Audit LLM benchmark arrivals, measurement clocks and outcome denominators before trusting tail latency. Includes a local HTTP experiment and acceptance gates.
Set a RAG answer, clarify and escalate boundary using authorised evidence, separate claim-support checks and explicit coverage and review-load denominators.
Ai2 releases AstaBrief for faster cited research reports. Check claim scope, checkpoint identity and the complete review workflow before an enterprise pilot.
Separate ONNX conversion, quantization and target-provider validation. Test tensor deviations and decision-threshold crossings before releasing an industrial edge model.
ServiceNow describes failure-driven synthetic training for agents. Audit verifier counterexamples, separate acceptance data and budget the complete pipeline before a pilot.
Bind AIOps approvals to object identity and observed state. Reject stale Kubernetes writes, test competing workers and reconcile ambiguous outcomes.
Google has announced Gemini 4 Argon with phased access and a larger output ceiling. Separate announced pricing from availability and bound the cost of long-running tasks.
Rehearse AI incident evidence preservation without delaying containment. Test release manifests, provenance, restricted originals and deliberately changed artifacts.
Test lower-precision KV state as a version-pinned capacity experiment. Separate cache memory, attention arithmetic, useful throughput and protected answer quality.
OpenShell separates runtime enforcement, credential isolation and policy proofs. Verify effective permissions and service-side outcomes before a read-only agent pilot.
Validate PDF cell, header, unit and footnote relationships before changing retrieval. Preserve structured provenance and quarantine unresolved table evidence.
Holo4 has different licenses for its 27B and 35B-A3B weights. Separate commercial rights, hosted terms and workflow evidence before choosing a deployment.
Stop turning incomplete maintenance follow-up into healthy labels. Build an event-or-censoring contract, audit replacement assumptions and choose supported evaluation horizons.
Retain rare failures without turning a biased trace sample into an SLO. A local Collector test separates request outcomes, late spans, restarts and routing changes.
Gemini Live Avatar adds video to enterprise voice agents. Calculate speaking-time charges, verify access and test fallback before enabling the visual persona.
An approval click does not prove effective review. Test recommendation-first and evidence-first interfaces with independently adjudicated cases and explicit error denominators.
Liquid AI releases a vision-language drafter. Check image-prefill latency, matched model artifacts, runtime restrictions and commercial licensing before rollout.
Decide when approved LoRA adapters can share a GPU. Test immutable routing, cold loads and working-set churn before consolidating serving pools.
Retrieved evidence can disappear inside the cross-encoder tokenizer. Inspect pair budgets, freeze candidates and test passage windows before replacing your embeddings.
OpenAI adds request-level prompt-cache diagnostics. Investigate missed reuse without replaying business actions, confusing comparison IDs with history or weakening controls.
A complete feature vector can conceal a disconnected sensor. Bound carry-forward by elapsed time, preserve provenance and compare availability with task error.
Transformers adds packed GGUF inference on Apple Silicon. Verify the kernel path, dequantization fallback and memory budget before replacing a working local runtime.
A restored checkpoint can replay an action that already committed elsewhere. Isolate recovery, reconcile remote effects and release only verified workflow operations.
Hugging Face Tokenizers v1 accelerates encoding, but its release candidate needs separate output, API and service-level checks before a production rollout.
A passing regression suite cannot settle an AI change review. Bind purpose, component deltas, supplier evidence and control tests to the exact candidate before promotion.
Claude Code fixes sandbox exclusion matching for compound Bash commands. Verify the running build, effective exceptions and mixed-chain behavior before approving rollout.
Chunked prefill protects running decodes, not automatically newly arrived short requests. Test per-class latency, token budgets and sequence limits before changing shared inference capacity.
Anthropic will embed Accenture evaluators inside its frontier-model work. Buyers still need evidence on access, reporting rights and the scope relevant to their own deployment.
Migrate embeddings as a coupled encoder-and-index route. Backfill authoritative text, capture updates and deletions, test shadow retrieval and keep rollback targets current.
Anthropic LSVP gives verified life-sciences teams broader model access, but requires 30-day retention. Separate eligibility, project scope and monitoring before a pilot.
Separate initial feature results, late corrections and online finalization. A Flink event-time walkthrough with explicit window boundaries, reconnect tests and a control-action boundary.
Define user-facing event counters, pair burn-rate windows and test alert recovery. A Prometheus walkthrough with six synthetic rule tests and explicit low-traffic safeguards.
Gemini 3.8 Live changes tool execution and conversation state. Separate audio interruption from backend cancellation before migrating write-capable voice agents.
Turn supplier instructions into deployment-bound operator scenarios. Test unsupported input, stale manuals, missing logs and safe stop before releasing AI capabilities.
Acceptance rate alone cannot justify a draft model. Compare drafting and verification overhead with committed-token savings under real serving load.
Compare direct multilingual retrieval and query translation on paired German questions and English manual evidence. Separate language gaps, truncation and ranking failures before release.
OpenAI’s new Data agent queries connected company data. Its recommended Sites publishing path copies analysis data and is unavailable in the EEA at launch. Separate query access from export approval.
Separate defect probability calibration from accept, review and reject decisions. A practical inspection policy with cost assumptions, lot-level evidence and review-capacity limits.
OpenAI Agents API supports self-hosted execution, but its managed sessions remain US-resident and ineligible for ZDR. Review the data boundary before connecting private tools.
Contain retry amplification across orchestrators, model gateways and tool clients with one deadline, bounded retry budgets and explicit side-effect reconciliation.
Connect deployment identity, approved purpose and registration evidence through a controlled handoff. Separate actor-specific routes, submission status and release reconciliation.
ChatGPT Images 2.5 introduces Flare and Sunburst. Compare verified access and pricing, then test unchanged product details and cost per approved asset.
Prefix caching avoids repeated prefill, not output generation. Evaluate cold and warm traffic, KV-memory pressure and cost per accepted response before claiming savings.
Design identifier-aware hybrid retrieval with exact lookup, lexical and vector candidates, RRF and query-class evaluation. Similar parts are not interchangeable.
OpenAI’s research disclosure reports growing agent use and continued human steering. Enterprise pilots should measure reviewed research outcomes, not runtime or token budgets.
WeatherNext 3 adds hourly forecasts and finer surface detail. Verify delivery latency, dataset access and conflicting licence documentation before an energy-planning pilot.
Allocate four IPv4 networks inside 10.42.0.0/22, expose a conflicting partner prefix and prepare a reviewable address plan with explicit capacity and routing assumptions.
Design asset-aware, forward-time backtests with causal features, mature labels and event-level acceptance criteria before deploying predictive maintenance.
NVIDIA has agreed to acquire Hugging Face and promises hardware choice. Enterprise buyers should verify artifact access, contract terms and recovery without live Hub dependencies.
Design OpenTelemetry Collector backpressure from an explicit outage and data-loss budget: align bounded sending queues, retry deadlines, persistent WAL storage, overload controls and fault-injection recovery tests.
An export-safe architecture for Article 50 transparency: provenance envelopes, transformation contracts, visible and machine-readable disclosures, and tests against metadata loss.
OpenAI’s GPT-6 Astra adds stronger computer use, a 1.05M-token context window and Critical cyber capability. Enterprise adoption needs a separate trust tier, cost gate and action boundary.
A production pattern for versioning prompts, datasets and graders together, running paired baseline/candidate evals in CI, and blocking regressions by protected segment before merge.
A production decision framework for choosing silo, pooled metadata filters, bridge routing or database-native tenant shards—and testing that foreign chunks never cross the retrieval boundary.
Apply ISA-18.2 alarm rationalization to predictive-maintenance model alerts: separate scores from episodes, tune thresholds with workload guardrails, stop chattering, and review bad actors.
A production framework for distilling large language models into lower-cost students while protecting task accuracy, safety segments, latency SLOs and rollback paths.
Turn RAG quality into a blocking CI decision with a versioned golden set, separate retrieval and generation metrics, protected-segment floors, baseline deltas and review bands.
A single MTTR number cannot show whether AIOps improved detection, diagnosis or repair. This measurement architecture defines lifecycle timestamps, stage metrics, guardrails and an auditable ROI evaluation.
A production validation loop for manufacturing digital twins: map ISO 23247 entities to sensor-quality, state-agreement and decision-fitness checkpoints; calibrate KS, PSI and residual thresholds; and separate recalibration from retraining.
Start with prompting, add retrieval for changing or traceable knowledge, and fine-tune only for stable behavioural defects. This decision framework compares volatility, latency, cost, governance exposure and measurable failure modes—and shows when RAG+PEFT is justified.
Across four independent evaluations, semantic chunking matched or lost to fixed-size and page-level baselines while costing up to 300 times more preprocessing time. What reproduces everywhere is that optimal chunk size depends on your corpus and your embedding model — so here is the recall@k evaluation to run before accepting any vendor default.
Correlation-based AIOps compresses alert storms into situations and leaves root cause as a ranked guess; causal engines traverse a dependency topology with fault-tree analysis to name one cause plus blast radius — but only while that topology is complete. A pipeline breakdown, decision table, failure modes and a replay evaluation you can run on your own incident history.
EU AI Act Article 72 requires a documented post-market monitoring system, but dashboards do not turn harm signals into validated change. A signal taxonomy, triage matrix, release-bundle attribution, fix-path table and incident-to-release-criterion ratchet for high-risk AI.
Turn AI literacy into an enforceable control: a role and use-case capability matrix, time-bound competence claims as identity attributes, IdP-backed entitlement, contextual acknowledgement, change-triggered recertification and aggregated learning telemetry.
A practical design for release gates that compare model candidates with a baseline by protected segment, block harmful regressions and route uncertain evidence to human review.
Build one versioned evidence platform for privacy and fundamental-rights assessment: shared system facts, separate DPIA/FRIA lenses, risk-to-control-to-test traceability, reassessment triggers and strict access separation for sensitive material.
Build production model routing as a versioned policy: named request classes, capability and quality floors, latency and cost budgets, bounded fallbacks, outage drills, redacted shadow traffic and invoice reconciliation.
A model version alone cannot explain a production decision. Bind the model and container digest, dataset and feature schema, sensor calibration, process envelope, thresholds and approvals into one release dossier—then test rollback of the compatible bundle before scaling.
AIOps cannot establish causality from incomplete or time-misaligned OT telemetry. Build a passive evidence path with asset and zone context, timestamp provenance, quality flags, retention tiers and readiness gates before automating correlation or remediation.
Persistent context, faster agent models and browser actions expand what AI can see and do. The enterprise response is a scoped control boundary—not ambient access.
A production architecture for proving which source data, transformations, features and retrieval-index versions shaped a high-impact AI release—then blocking promotion when contracts, cohort evidence or approvals are stale.
Quantisation and adapters can reduce memory and training cost, but production promotion needs evidence. Compare baseline and candidate variants across German, long-context, numeric and structured-output segments, then test TTFT, throughput, runtime compatibility and rollback under controlled load.
Google, OpenAI and Anthropic are bringing scheduled AI work into their products. For European enterprises, the key question is not when a task runs, but who controls its data, authority and evidence trail.
A production control-plane pattern for governing hosted models, open weights, retrieval components and fine-tunes: approved tuples, region and endpoint gates, component bills of materials, evidence and supplier-change re-review.
Choose device edge, plant edge or cloud by failure domain, data gravity and recovery objectives—not by slogans. A practical partition-tolerant architecture for industrial AI.
OpenAI is expanding free ChatGPT access. For European employers, the practical response is a managed workspace, explicit data boundaries and tested controls—not a blanket ban or ungoverned personal use.
A practical release-control pattern for LLM systems: an immutable bundle that pins model, prompt, tokenizer, evaluation evidence, approval and rollback target—so teams can reproduce and reverse a production change.
Industrial AI needs a versioned data contract before model selection: stable asset identity, semantic meaning, units, ranges, timestamps, quality, machine state, ownership and permitted use—validated at the edge and represented through AAS.
Run self-hosted LLM inference against TTFT, inter-token latency, queue time, end-to-end latency and error budgets. Use GPU telemetry to explain SLO failures, then validate capacity with a controlled saturation matrix.
OpenAI’s GPT-5.6 preview and Alibaba Cloud’s Qwen3.8-Max announcement make a familiar enterprise problem more urgent: evaluate the deployed workload, data path and update control—not a model name alone.
Build a versioned AI intake that maps each use case to its model, supplier, data, process and affected groups—then use policy-as-code to assign controls and block incomplete procurement or deployment.
Connected work agents can access data, tools and actions. The practical enterprise question is how to constrain, evaluate and revoke that access before a useful pilot becomes an uncontrolled production dependency.
A production architecture for industrial AI: separate observation, inference and safety/control planes; use an OT DMZ and least-privilege identities; keep AI out of direct PLC and SIS command paths.
Google’s Gemini Notebook expands a source-grounded research workspace across connected Gemini surfaces and adds native code execution for eligible plans. European enterprises should treat each notebook as a governed integration boundary: control sources, permissions, evidence and data-processing scope before scaling it.
How to turn structured LLM responses into versioned, validated API contracts: JSON Schema, semantic rules, typed refusal paths, transformation layers and operational metrics.
AI-assisted AI development is accelerating, but autonomous self-upgrading remains unproven. European enterprises should treat every material model or agent change as a controlled, testable and reversible release.
Enterprise RAG security requires more than ACL filters. Build an operating model that proves identity, authorization, redaction, retrieval decisions, revocation and incident evidence for every answer.
OpenAI’s Astra research, agent-boundary failures and a call to pace frontier AI point to one enterprise priority: expand controls before expanding autonomy.
A production architecture for RAG prompt-injection resistance: separate data from instructions, authorize every tool action, preserve provenance, and test poisoned documents before deployment.
