Code 231, Kenitra (2026)

04/10/2026

Enterprise AI gets complex fast.

Once AI moves beyond a simple chatbot, the real work begins: controlling cost, grounding outputs, connecting systems, evaluating quality, orchestrating agents, and deciding when humans need to step in.

Here are 15 concepts every AI leader should understand:

→ Tokenomics - track and budget token usage across teams and models.
→ Ontology - define business entities, relationships, and shared meaning.
→ Semantic Layer - create governed metrics for agents and BI tools.
→ Agent Harness - wrap models with tools, memory, guardrails, and runtime logic.
→ Context Engineering - control what information enters the model context.
→ GraphRAG - combine vector, keyword, and graph retrieval for stronger grounding.
→ Model Routing - choose models based on quality, cost, complexity, and latency.
→ AI Gateway - centralize auth, quotas, caching, fallbacks, and logging.
→ MCP - connect agents with tools, data, and enterprise systems.
→ Guardrails - validate prompts and outputs against policy and risk.
→ Evals - measure model behavior before changes reach production.
→ LLM Observability - trace prompts, tools, tokens, failures, and cost.
→ Agent Orchestration - coordinate multiple agents across multi-step workflows.
→ Human-in-the-Loop - escalate high-risk or low-confidence actions.
→ Semantic Caching - reuse similar responses to reduce latency and spend.

The key shift is simple:

Enterprise AI is no longer just about choosing the best model.

It is about building the system around the model so AI becomes reliable, governed, observable, and scalable.

Which of these 15 concepts are you currently working on?

04/10/2026

Your AI budget deserves more than a surface-level grasp of LLMs.

__________

Use this nine-topic map to improve build, buy and review decisions.
Bring one real work task to each session.
Write down what you can now explain, test or challenge.
Watching alone will not establish mastery.

1 → Transformers
Trace how tokens and attention shape an answer.
Explain that flow to a colleague in plain words.
Use a short input so you can inspect each step.
lnkd .in/dGnQW39t

2 → Transformer Tricks
Look for the choices that affect efficiency.
Ask which tradeoffs matter for your team's workload.
Test the speed and quality you actually need.
lnkd .in/dT_VEpVH

3 → Large Language Models
Connect model architecture to output behaviour.
Keep a record of where the model fails your task.
Use the same examples when you compare models.
lnkd .in/dwjjpjaP

4 → LLM Training
Separate pretraining from task adaptation.
Learn where SFT and LoRA fit before funding a build.
Ask what data you have the right to use.
lnkd .in/dSi_xCEN

5 → LLM Tuning
Learn what RLHF, PPO and DPO try to change.
Define the behaviour you want in concrete terms.
Include examples of answers you would reject.
lnkd .in/dUK5djpB

6 → LLM Reasoning
Study RL scaling and GRPO with a hard task in mind.
Check the final result against an independent answer.
A long explanation still needs evidence.
lnkd .in/dAGQTNAM

7 → Agentic LLMs
Map RAG, tool calls and agents to one workflow.
Name the tools and data the system can access.
Set an approval point before it changes anything important.
lnkd .in/dWD4j7vm

8 → LLM Evaluation
Build a small set of realistic cases with known outcomes.
Compare LLM-as-a-Judge scores with human review.
Keep failure examples when you update the system.
lnkd .in/ddxE5zvb

9 → Applied LLM Systems
Use this source theme to plan deployment and monitoring.
Give the workflow an owner and a rollback route.
Stanford's official lecture 9 is Current trends.
lnkd. in/dGsTd8jN

Resource note: source links 1 and 9 open the same video.
Use the official CME 295 index to navigate the lectures:
https://lnkd .in/eJ-H8YnY

Choose one lecture and one real business question.
Turn the lesson into a small test your team can repeat.
Use what you learn to challenge the next AI proposal.

Which topic would most improve your next build-or-buy decision?

03/10/2026

𝐌𝐨𝐬𝐭 𝐨𝐫𝐠𝐚𝐧𝐢𝐬𝐚𝐭𝐢𝐨𝐧𝐬 𝐬𝐭𝐢𝐥𝐥 𝐦𝐞𝐚𝐬𝐮𝐫𝐞 𝐀𝐈 𝐬𝐮𝐜𝐜𝐞𝐬𝐬 𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐞𝐱𝐩𝐞𝐫𝐢𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 𝐚𝐜𝐭𝐢𝐯𝐢𝐭𝐲.

The stronger organisations measure AI through business leverage.

That is the real shift happening in 2026.

AI ROI is no longer about proving that AI works.

That phase is over.

The real challenge now is building AI systems that improve operational efficiency, reduce cost exposure, accelerate ex*****on, and create measurable business impact at scale.

The organisations moving ahead understand something important:

AI value compounds only when economics, operations, and governance are aligned together.

𝐖𝐡𝐢𝐜𝐡 𝐢𝐬 𝐰𝐡𝐲 𝐦𝐚𝐭𝐮𝐫𝐞 𝐀𝐈 𝐥𝐞𝐚𝐝𝐞𝐫𝐬 𝐚𝐫𝐞 𝐟𝐨𝐜𝐮𝐬𝐢𝐧𝐠 𝐨𝐧:
→ Sustainable cost optimisation
→ Workflow-level efficiency gains
→ Revenue acceleration opportunities
→ Risk-aware AI operations
→ Long-term operational scalability

Because enterprise AI is no longer an innovation showcase.

It is becoming part of core business infrastructure.

And infrastructure must justify itself continuously.

The companies creating durable AI advantage are not simply deploying more AI tools.

They are building operational systems where AI contributes measurable strategic value across the organisation.

That changes how leadership evaluates success.

The future winners will not necessarily have the highest AI adoption.

They will have the strongest AI operating economics.

P.S. Many companies still ask:

“How much AI are we using?”

The more mature organisations ask:

“How much business leverage are we creating from every AI investment?”

03/10/2026

How Your Code Really Controls Hardware

03/10/2026

AI mastery does not require a technical background.

You do not need to start with coding, machine learning mathematics, or complex architectures.

Start with understanding how modern AI tools actually work together.

Here’s a practical roadmap:
→ AI Basics
Learn LLMs, prompts, context, hallucinations, and multimodal AI.

→ AI Assistants
Understand where tools like ChatGPT, Claude, Gemini, Perplexity, and NotebookLM fit best.

→ Prompting
Give AI a clear task, useful context, constraints, examples, and the output format you want.

→ AI Capabilities
Use web search, deep research, data analysis, image understanding, content creation, and workspaces.

→ Personalization
Set memory, custom instructions, projects, file uploads, and trusted knowledge sources.

→ Connectors & MCP
Connect AI to approved data, apps, tools, and actions while controlling permissions.

→ Skills & Customization
Use templates, plugins, reusable skills, SKILL.md files, and custom workflows.

→ Agents & Automation
Learn how agents use tools, workflows, triggers, human approval, and scheduled tasks.

→ AI Best Practices
Give context, use specific prompts, verify important outputs, limit access, reuse workflows, and keep humans in control.

The biggest mistake is trying to learn everything at once.

Start with one AI assistant.

Learn prompting.

Add your files and context.

Then connect tools, build reusable workflows, and gradually move into agents and automation.

That is a much simpler path from AI beginner to confident AI user.

03/10/2026

Your agent roadmap needs an operating model before it needs more agents.

__________

The next step is getting agents to work reliably inside a real business.
Use these six trends to audit one workflow already on your roadmap.
Start with the work, the owner and the limits.
Then choose the tools.

1. Multi-agent orchestration
Split research, retrieval and review when separate roles help.
Give each agent a clear handoff and test the combined result.
Examples: LangGraph, AutoGen, AWS Bedrock, Relevance AI and CrewAI.
Add agents only when the extra coordination earns its cost.

2. Context engineering
Decide which records, instructions and past decisions each step needs.
Test retrieval quality and keep irrelevant context out.
Examples: mem0, LlamaIndex, Pinecone and Zep.
More memory does not automatically mean better decisions.

3. Guardian agents
Check policy, tool permissions and risky outputs during ex*****on.
Define when to block an action or hand it to a person.
Examples: Lakera, Salesforce AITM, Patronus AI and Arthur.
Keep a named human owner for the controls themselves.

4. Agentic commerce
Map search, comparison, approval and purchase as separate steps.
Set spending limits before connecting a payment tool.
Examples: Amazon Rufus, Stripe Agent Toolkit, Walmart Sparky,
OpenAI Operator and Perplexity Shop.
Use transaction logs to track what was approved and bought.

5. Coding agents at scale
Claude Code, GitHub Copilot, Cursor, Replit Agent, Devin and Windsurf
can support a workflow from codebase analysis to pull requests.
Require tests, review and a clear deployment decision.
Measure accepted changes and defects alongside time saved.
Fast code creation still leaves someone responsible for shipping it.

6. Governance and evals
Define success cases, failure cases and approval rules before release.
Use LangSmith, Braintrust, Weights & Biases or promptfoo
to support tracing, evaluation and production review.
Check quality, cost and failure recovery after each meaningful change.

Treat the graphic as a practical watchlist.
Its Adobe shopping and McKinsey adoption figures are 2025 snapshots.
Gartner's inquiry growth measures interest, not deployment success.
The 2030 guardian-agent and commerce figures are forecasts.

Tool names reflect the source examples, including legacy names.
Check current support, availability and fit before choosing a stack.
The useful question is which part of your workflow needs improvement.
A larger collection of agents does not settle that question.

Pick one process with a clear owner.
Define its context, permissions and handoffs.
Decide what good output looks like and how to recover from failure.
Scale only after the results justify the next step.

Which of these six areas is holding back your agent rollout?

03/10/2026

Your RAG system can retrieve the wrong evidence beautifully.

__________

Before you swap models, inspect the retrieval path.
The architecture should fit the questions your team needs to answer.
These five patterns give you a practical way to choose what to test.

→ 1. Hybrid RAG: match meaning and exact terms
Combine dense vector retrieval with sparse keyword search such as BM25.
Merge the ranked results with Reciprocal Rank Fusion, then select the context for the LLM.
Test this when product codes, names or policy terms get lost in semantic search.
Check the retrieved chunks before judging the final answer.

→ 2. GraphRAG: follow relationships
Represent entities and their connections in a knowledge graph.
Use relevant subgraphs or community summaries to supply context.
Try it when a question spans people, projects, places or dependencies.
Inspect entity extraction and links: a graph can preserve errors too.

→ 3. Agentic RAG: plan the retrieval
Let a planner choose among vector search, web search and SQL tools.
Use returned evidence to decide the next step.
Set tool permissions, a step limit and a stopping rule before deployment.
Confidence alone is not proof that the answer is supported.

→ 4. Corrective RAG: check weak retrieval
Evaluate retrieved evidence before using it to generate an answer.
Route weak results to another retrieval action, such as query rewriting or a web fallback.
Test the fallback as well as the main route.
A relevance grade does not guarantee factual accuracy.

→ 5. Multimodal RAG: preserve visual evidence
Retrieve from text, images, charts and tables when layout carries meaning.
CLIP and ColPali are examples of different retrieval approaches, not interchangeable components.
Check the model, representation and index together.
Test a question that needs the chart, not just its nearby text.

The diagrams are simplified patterns.
They combine some indexing and query stages for clarity.
GraphRAG and CRAG implementations vary from these sketches.
“Top 5” is an editorial selection, not a benchmark ranking.

Run one useful comparison before adding complexity.
Build a small set of real questions and expected supporting sources.
Include exact lookups, cross-document questions and missing evidence.
Compare retrieved evidence, answer support, latency and cost on the same set.
Keep a simple baseline so you can see what the added design actually changes.

Name the failure you need to fix.
Choose the pattern that addresses it.
Measure the result before expanding the system.

Which fails most in your current RAG setup: exact terms, relationships, planning, validation or visual evidence?

03/10/2026

Repeated AI rework often starts with missing context.

__________

The three levels are an illustrative framework.
Use them to decide how much structure a task deserves.
The deeper work is defining the job, the context and the checks.
More setup only helps when it improves the result.

Level 1 → Get an answer
Ask a clear question for a small, low-risk task.
Give one sample to show a style.
Give a few samples when the pattern matters.
Ask for options, then choose the one that fits.

Quick edits, rewrites, shortening and translation can start here.
Read the result before you use it.
For a simple one-off job, this may be enough.
Do not build a complex workflow just to fix one sentence.

Level 2 → Direct the work
Define the role and the exact job.
Name the reader, the goal and what to leave out.
Choose the response format: bullets, a table or paragraphs.
Show a finished example so quality has a visible standard.

State the hard rules and the point at which it should ask first.
Attach the files that matter.
Enable web or live data access only when the task needs it.
Then check whether the answer follows the brief.

Try this on a weekly client update.
Provide the audience, current facts, desired length and last week's format.
Ask for unresolved questions to be flagged.
Review the draft against those inputs before sending it.

Level 3 → Make good work reusable
Choose a model that fits the task, cost and quality target.
Ask for checks before a final answer, then verify the result yourself.
For work with several steps, review a plan before ex*****on.
State the problem and constraints without dictating every step of the fix.

Save stable instructions and package repeated work into a Skill.
Keep relevant client files and rules together in a Project.
The graphic uses Claude and Claude Code examples, including CLAUDE.md.
Those features are implementation choices; support varies by product.

MCP can connect supported tools and data sources.
Gmail, Notion, Slack and Drive are examples in the source.
Connections need configuration, access and clear permissions.
They do not remove your duty to check consequential actions.

Combine role, context, examples and iteration.
Test the workflow on real inputs and keep the failures.
When a rule or example stops helping, revise it.
Reusable instructions still need maintenance and judgment.

Pick one recurring task and record the time spent reviewing it.
Add only the missing context, example or check.
Compare the next result with your current process.
Keep the changes that make the output easier to trust and use.

Which recurring task needs a better brief before it needs a better model?

03/10/2026

AI governance breaks when ownership is unclear.

The challenge is not lack of policies. It is getting the right leaders to own the right decisions across strategy, technology, delivery, and security.

The CEO sets the direction.
They define risk appetite, business outcomes, investment priorities, ethical expectations, and the level of accountability the board expects.

The CIO connects AI to the enterprise.
They align data, processes, platforms, vendors, architecture, and operating models so AI becomes part of how the business actually runs.

The CTO turns strategy into working systems.
They own model integration, engineering standards, scalability, platform design, deployment choices, and the technical decisions required to move AI from pilot to production.

The CISO makes sure that progress remains trusted.
They focus on security, privacy, IAM, monitoring, compliance, incident response, prompt injection, data leakage, and resilience.

The real value sits in the overlaps.

CEO + CIO = strategic alignment.
CEO + CTO = risk and trust.
CIO + CISO = policy translated into operational controls.
CTO + CISO = secure engineering.
CIO + CTO = scalable platform delivery.

Strong AI governance is not one executive approving everything. It is a shared operating model with clear ownership, measurable accountability, and controls built into delivery.

That is how organizations move faster without losing trust.

Which leadership overlap is the hardest to get right in your organization?

02/10/2026

Most people use ChatGPT like Google. Type a few words, get a generic answer, close the tab.

Then they say "AI isn't that useful."

The tool isn't the problem. The prompt is.

Here's the 6 step habit that fixes most bad outputs:

1. Set the right role.
"Act as a marketing strategist" changes the tone, depth, and angle of the whole answer.

Without it, you get the average of everything.

2. Give clear context.
Who is it for? What's the goal? The more background you share, the less guessing it does.

3. Ask for a specific output.
"Create a 7-day content plan in a table with captions and hashtags" beats "give me content ideas" every time.

Format, length, and tone, say all three.

4. Refine and improve.
The first answer is a draft, not the final. Try "make it shorter," "make it more conversational," or "give me a stronger opening line."

5. Use the key features.
Browse the web for live info.
Upload files to summarize or analyze.
Generate images.
Set custom instructions so you stop repeating yourself.

6. Save and organize.
This is the one almost everyone skips.
When a prompt works, keep it.
A small library of content ideas, email templates, blog outlines, and research prompts will save you hours every week.

The biggest shift is this: stop treating prompts as one off questions.
Start treating them as reusable assets.

The people getting great results from AI aren't smarter.
They're just clearer about what they want, and they don't start from zero each time.

Which step are you guilty of skipping?

Be honest, I'll go first: it's #6 for most of us.

---

Code 231

04/10/2026

04/10/2026

03/10/2026

03/10/2026

03/10/2026

03/10/2026

03/10/2026

03/10/2026

03/10/2026

02/10/2026

Address

Website

Alerts

Shortcuts

Share

Category

Other Information Technology Companies in Kenitra

Listing takedown request