# Brahim Bousnguar > Tech lead and senior e-commerce integration consultant with 9+ years of experience specializing in SAP Commerce Cloud (Hybris/CCv2), MuleSoft Anypoint Platform, and Salesforce integration. Based in France, working with enterprise clients across 15+ countries. Bilingual English/French. Key facts: - Current role: MuleSoft Integration Specialist at Roche Bobois, the French luxury furniture retailer (since January 2025) — API-led integration architecture, DataWeave, Salesforce CRM data flows. - Previous: Senior SAP Commerce Cloud Consultant at Miele (2022–2024, cloud-native CCv2 migration, plus a first SAP BTP Kyma proof of concept); Technical Lead at BYK, the speciality-chemicals group (2018–2021, multi-country B2B platform, Hybris 6.7 → CCv2); SAP Hybris Integration Specialist at SEB Group (2016–2018, multi-brand platform). - Certifications: GitHub Copilot (GH-300) (Sep 2026); SAP Certified — Solution Architect — SAP Customer Experience (Mar 2026); SAP Certified — Positioning SAP Customer Experience Solutions (Mar 2026); SAP Certified — Business User — SAP Commerce Cloud (Mar 2025); SAP Certified Development Professional — SAP Commerce Cloud Developer (2023); SAP Certified Associate — Hybris Commerce Business Analyst 6.3 (2018). Currently pursuing MuleSoft Anypoint Platform certifications. The 2026 and 2025 SAP certifications are verifiable on Credly: https://www.credly.com/users/brahim-bousnguar. The GitHub Copilot certification is verifiable on Microsoft Learn: https://learn.microsoft.com/api/credentials/share/en-us/BrahimBousnguar-7690/20B68101386B5A?sharingId=9C9C9C1DC7CAD24F - AI-augmented workflow: daily use of GitHub Copilot, Ollama-based LLM CLI tooling, Claude, and OpenAI Codex CLI. Side projects around MCP servers, local LLM benchmarking, and AI coding-assistant cost tracking. - Education: State Engineer in Computer Science & Business Intelligence, INSEA Rabat (2015). - Availability (updated 2026-09-26): open to a permanent (CDI) tech-lead or integration-architect role on MuleSoft and SAP Commerce Cloud; can start three months after an offer (notice period); remote, or on site in Nantes, France. Details: https://heybrahim.com/now.html ## Pages - [Portfolio home](https://heybrahim.com/): hero, integration diagram (SAP Commerce Cloud → MuleSoft → Salesforce/ERP), client case studies (Problem/Built/Result), side projects and open source, experience, stack, certifications, contact. English; the French version is at https://heybrahim.com/fr/. - [About](https://heybrahim.com/about.html): career story, timeline, current focus and what comes next. - [Now](https://heybrahim.com/now.html): what Brahim is working on this month and his availability (dated). - [Talks](https://heybrahim.com/talks/): all slide decks. [The agentic software factory](https://heybrahim.com/talks/software-factory/) (English; French at https://heybrahim.com/talks/software-factory/fr/; about 20 minutes, built on the software-factory note; ready for meetups, not yet given), [MCP: giving your agents hands](https://heybrahim.com/talks/mcp/) (English; French at https://heybrahim.com/talks/mcp/fr/; open-source demo server at https://github.com/brbousnguar/mcp-staffing-demo), [An agent fleet at home](https://heybrahim.com/talks/openclaw/) (English; French at https://heybrahim.com/talks/openclaw/fr/; 17 agents on a Mac mini). Listed on the About page under Talks. - [How I work with AI](https://heybrahim.com/ai.html): his agent workflow (issue → agent on its own branch → proof in a real browser → independent model review → a PR he merges), the tools he built for it and the guardrails. - [AI Brief](https://ai.heybrahim.com/): the archive of his daily AI-news brief, on its own host since 2026-10: every brief with its sources and an audio version, and running stories tracked across days. English. - French versions: [Accueil](https://heybrahim.com/fr/), [À propos](https://heybrahim.com/fr/a-propos.html), [Maintenant](https://heybrahim.com/fr/maintenant.html), [IA](https://heybrahim.com/fr/ia.html). ## Machine-readable - [MCP server](https://api.heybrahim.com/mcp): this portfolio as a remote MCP server (Streamable HTTP, no key). Tools: get_profile, get_availability, list_items, get_item, search, request_cv (files a CV request like the site form; Brahim answers by hand). - [Agent discovery](https://heybrahim.com/.well-known/ard.json): ARD manifest listing the MCP server (same file at /.well-known/ai-catalog.json). - [MCP server card](https://heybrahim.com/.well-known/mcp.json): the MCP server's card (same file at /.well-known/mcp/server-card.json), endpoint https://api.heybrahim.com/mcp. - [profile.json](https://heybrahim.com/profile.json): JSON Resume-style structured profile (basics, work history, skills, projects, education). - [CV on request](https://heybrahim.com/cv.html): the full CV names clients, so it is sent by email on request (fill in the form; Brahim approves and sends it himself). The public facts are in profile.json and on this site. ## Learn Interactive lessons (English), in track order: each page has live diagrams of the platform at work (in the plain text, each diagram is one sentence) and a three-question check. Hub: https://heybrahim.com/learn/ · MuleSoft track: https://heybrahim.com/learn/mulesoft/. Markdown at the same URL with `.md`. - [Mule apps and flows](https://heybrahim.com/learn/mulesoft/mule-app-flows.html) (2026-10-21): What a Mule app is made of, how one Mule event moves through a flow, what On Error Propagate and On Error Continue send back to the client, and how Flow Reference, Choice, Scatter-Gather and For Each change the path. Five live diagrams and a three-question check. Markdown: https://heybrahim.com/learn/mulesoft/mule-app-flows.md - [Watermarks](https://heybrahim.com/learn/mulesoft/watermarks.html) (2026-10-21): How a polling Mule flow remembers where it stopped, so the next poll picks up only what is new. The Database source that keeps the bookmark for you, the Object Store pattern for everything else, and what a redeploy or a second replica does to it. Four live diagrams and a three-question check. Markdown: https://heybrahim.com/learn/mulesoft/watermarks.md - [Queues: VM and Anypoint MQ](https://heybrahim.com/learn/mulesoft/queues.html) (2026-10-21): How a Mule flow hands work to a queue with the VM connector inside one app, or with Anypoint MQ between apps, and what happens when the consumer fails, crashes or runs slower than its lock. Four live diagrams and a three-question check. Markdown: https://heybrahim.com/learn/mulesoft/queues.md ## Client work Case studies (English; French versions under https://heybrahim.com/fr/missions/). Markdown at the same URL with `.md`. - [API-led integration for Roche Bobois](https://heybrahim.com/work/api-led-luxury-retail.html) (2026-09-18): How I moved Roche Bobois's e-commerce, ERP, CRM and third-party flows from point-to-point links to reusable API-led integration on MuleSoft Anypoint, with DataWeave and Salesforce CRM flows. Ongoing since 2025. Markdown: https://heybrahim.com/work/api-led-luxury-retail.md - [Moving Miele to SAP Commerce Cloud](https://heybrahim.com/work/cloud-commerce-migration-appliances.html) (2026-09-18): How I helped move Miele's global B2C platform from on-premise SAP Hybris to SAP Commerce Cloud (CCv2) with zero downtime, optimised OCC APIs, started a proof of concept on SAP BTP Kyma and set up Jenkins CI/CD. 2022–2024. Markdown: https://heybrahim.com/work/cloud-commerce-migration-appliances.md - [One B2B commerce platform for 15+ countries](https://heybrahim.com/work/multi-country-b2b-platform.html) (2026-09-18): How I led BYK's B2B portal for 15+ countries: one SAP Commerce Cloud platform with SAP ERP, logistics and payment integrations and custom OCC APIs. Markdown: https://heybrahim.com/work/multi-country-b2b-platform.md ## Projects Side projects, one page each (English; French versions under https://heybrahim.com/fr/projets/). Markdown at the same URL with `.md`. - [Where's my sofa?](https://heybrahim.com/projects/wheres-my-sofa.html) (2026-10-18): Where's my sofa? is a customer care agent for a made-to-order furniture shop, built on MuleSoft Agent Fabric: an Agent Script broker behind Omni Gateway, a Mule MCP server and a Mule System API that holds the rules. Markdown: https://heybrahim.com/projects/wheres-my-sofa.md - [tabkit](https://heybrahim.com/projects/tabkit.html) (2026-10-14): tabkit is an open-source, read-only MCP server with a Chrome extension that lets Claude Code, GitHub Copilot or Codex read Jira and Confluence with your logged-in Chrome session, without an Atlassian API token. Markdown: https://heybrahim.com/projects/tabkit.md - [pilotkit](https://heybrahim.com/projects/pilotkit.html) (2026-10-12): pilotkit is an open-source CLI that connects a GitHub Copilot seat to Jira and Confluence with one command, through a read-only MCP server, and adds a local desk that shows the AI credits every answer costs. Markdown: https://heybrahim.com/projects/pilotkit.md - [mulewatch](https://heybrahim.com/projects/mulewatch.html) (2026-09-08): mulewatch is an open-source, read-only MCP server that lets Claude, Cursor or any MCP client read live and archived MuleSoft Anypoint logs, list deployed apps, API Manager instances and Exchange assets. Markdown: https://heybrahim.com/projects/mulewatch.md - [vibetidy](https://heybrahim.com/projects/vibetidy.html) (2026-09-08): vibetidy is a zero-dependency Node CLI that writes a README from what a repository actually contains, with an LLM that's only allowed to format verified facts, and warns when a feature-sized commit has no issue behind it. Markdown: https://heybrahim.com/projects/vibetidy.md - [gemmaquiz](https://heybrahim.com/projects/gemmaquiz.html) (2026-08-12): gemmaquiz turns a rough subject into a multiple-choice quiz: a local Gemma model on Ollama cleans up the topic, reads the real Wikipedia article and writes the questions. No cloud LLM, no API keys. Markdown: https://heybrahim.com/projects/gemmaquiz.md - [Meterlex](https://heybrahim.com/projects/meterlex.html) (2026-07-30): Meterlex reads the session logs of Claude Code, Codex, Gemini CLI, Antigravity, Ollama and Copilot CLI, prices every model call at the public API rate and compares it with what your subscriptions cost. Local-first, token counts only. Markdown: https://heybrahim.com/projects/meterlex.md - [FocusTracker](https://heybrahim.com/projects/focus-tracker.html) (2026-07-15): FocusTracker is a native macOS focus timer with a menu-bar countdown, categories, session notes and a monthly dashboard. Sessions stay in a local file on your Mac or go to your own Supabase database. Markdown: https://heybrahim.com/projects/focus-tracker.md ## Notes Technical notes (English), newest first. Plain-text Markdown of each is at the same URL with `.md`; everything in one file: https://heybrahim.com/llms-full.txt. French versions of the notes are listed at https://heybrahim.com/fr/notes/ (Atom: https://heybrahim.com/fr/notes/feed.xml), each with its `.md` copy. ### Articles - [The Ralph loop: how a bash one-liner moved into the model](https://heybrahim.com/notes/ralph-loop-moved-into-the-model.html) (2026-11-11): Coding agents used to stop before the work was done, so people wrapped them in a bash loop that restarted them on the same prompt. That trick, the Ralph loop, then moved into the tools as a goal command, and then into the models. How it works, and how Claude Code and Codex handle it today. Markdown: https://heybrahim.com/notes/ralph-loop-moved-into-the-model.md - [Levi's rebuilt its development process before it added agents](https://heybrahim.com/notes/levi-strauss-rebuilt-the-process-before-the-agents.html) (2026-11-09): Levi Strauss & Co. cut the build of a store app from nine months to two and a half. The agents were not the first step. The team first redrew the whole process, split what agents do from what people decide, then built 17 agents on GitHub Copilot to fit it. Markdown: https://heybrahim.com/notes/levi-strauss-rebuilt-the-process-before-the-agents.md - [Personal AGI is an agent you own, not a chat you rent](https://heybrahim.com/notes/personal-agi-own-not-rent.html) (2026-11-09): Garry Tan says AGI won't arrive as one big announcement. It's already here, spread out, as an agent that runs on your own machine and works from your own notes. What he means by personal AGI, how he builds his, and why he says you should own your skill files. Markdown: https://heybrahim.com/notes/personal-agi-own-not-rent.md - [The abundance of intelligence: what changes when anyone can buy it](https://heybrahim.com/notes/abundance-of-intelligence.html) (2026-11-06): Expert thinking used to be scarce, sold by the hour and rationed by price. Now anyone can buy it for a few dollars a month, and its price falls about ten times a year. What the research says this changes for work, school and our own thinking, and what stays scarce. Markdown: https://heybrahim.com/notes/abundance-of-intelligence.md - [Two quarters to stop writing code by hand](https://heybrahim.com/notes/two-quarters-to-stop-writing-code.html) (2026-11-06): Datadog's co-founder told the whole engineering team that within two quarters, they would stop writing code by hand. What CEO Olivier Pomel said about that deadline, why Datadog still isn't "token maxing", and what it changes for developers. Markdown: https://heybrahim.com/notes/two-quarters-to-stop-writing-code.md - [Ask what the green check proved](https://heybrahim.com/notes/ask-what-the-green-check-proved.html) (2026-11-04): When a check goes green, ask what it proved, then what it didn't. Addy Osmani's checklist for critical thinking when agents do the work, held up against my own setup, where agents merge and deploy about 80 pull requests a day. Markdown: https://heybrahim.com/notes/ask-what-the-green-check-proved.md - [Before "thinking models" — the dark ages (2022–2024)](https://heybrahim.com/notes/before-thinking-models-dark-ages.html) (2026-11-04): From ChatGPT's launch in November 2022 to September 2024, chat models answered in one pass, with no step to check their work. How thinking models arrived, what they fixed and what they didn't, and what thinking costs in my own apps. Markdown: https://heybrahim.com/notes/before-thinking-models-dark-ages.md - [Intelligence on tap: build today what everyone gets later](https://heybrahim.com/notes/intelligence-on-tap-consumer-ai.html) (2026-11-04): An AI app that costs $1,000 a month is enterprise software today, and the same app becomes a consumer product once AI gets a thousand times cheaper. What I took from Garry Tan and Mark Pincus, the founder of Zynga, on building it now. Markdown: https://heybrahim.com/notes/intelligence-on-tap-consumer-ai.md - [Spending more tokens only pays with good skills and real tests](https://heybrahim.com/notes/tokenmaxxing-fat-skills-real-tests.html) (2026-11-01): Spending more tokens pays only if skills hold your judgement and tests catch the bad output. What I took from Garry Tan's episode on tokenmaxxing, held against my own skills, hooks and token bill. Markdown: https://heybrahim.com/notes/tokenmaxxing-fat-skills-real-tests.md - [Agents replace the form, not the data](https://heybrahim.com/notes/agents-replace-the-form-not-the-data.html) (2026-10-29): Peter Steinberger, who built OpenClaw, told Y Combinator that 80% of apps will go away, because an agent on your own computer manages data better than an app. In my fifty apps, the agent took over the typing, but the data still lives in the app, where I can check it. Markdown: https://heybrahim.com/notes/agents-replace-the-form-not-the-data.md - [Sign the limits, not the purchase](https://heybrahim.com/notes/sign-the-limits-not-the-purchase.html) (2026-10-28): AP2 is named for payments, but Alan from Google calls it delegated user authority. You sign an agent's limits once, on a screen no agent controls, and it acts inside them later without you. Markdown: https://heybrahim.com/notes/sign-the-limits-not-the-purchase.md - [Keep the agent session next to the code](https://heybrahim.com/notes/agent-session-next-to-the-code.html) (2026-10-27): An agent session holds the reasons behind a change, so Thomas Dohmke wants it kept in Git next to the commit it produced. His company Entire builds that link, and my own sessions are deleted long before the pull requests they wrote. Markdown: https://heybrahim.com/notes/agent-session-next-to-the-code.md - [An agent can carry your request, not your authority](https://heybrahim.com/notes/agents-carry-requests-not-authority.html) (2026-10-27): When your agent asks someone else's agent for something, only that agent's own owner can approve an irreversible step. That's the rule under DotSpace, the network of shared rooms that Chi Wang of AG2 showed live at AGNTCon + MCPCon, where 339 agents of 105 owners posted in three weeks. Markdown: https://heybrahim.com/notes/agents-carry-requests-not-authority.md - [Agents don't fit Kubernetes, and three more problems still open](https://heybrahim.com/notes/agents-dont-fit-kubernetes-open-problems.html) (2026-10-26): Katelyn Lesse, head of platform at Anthropic, lists four problems to solve before multiagent systems can do serious work, and I check my own 17 agents against each. Agents don't fit Kubernetes, act in a human's name, can pass an attack on to other agents, and need a kill switch that stops them all. Markdown: https://heybrahim.com/notes/agents-dont-fit-kubernetes-open-problems.md - [How the Claude Code team reviews code that Claude writes](https://heybrahim.com/notes/claude-code-team-code-review.html) (2026-10-26): Claude writes about 90% of Claude Code, and the team still reviews every change. Machines now do most of the reading: six layers of checks, a review level set by the risk of each change, and a cap on rounds so that review ends. Markdown: https://heybrahim.com/notes/claude-code-team-code-review.md - [Every screen in this SaaS app is an MCP App](https://heybrahim.com/notes/every-screen-is-an-mcp-app.html) (2026-10-26): Mathew Goldsborough of NimbleBrain showed a SaaS-style task app where every panel is an MCP App, put together at runtime by an open-source host. I read its code and docs to see how an app finds its place, shows a count without the model, hears it was installed, and shares one theme. Markdown: https://heybrahim.com/notes/every-screen-is-an-mcp-app.md - [A sandbox inside the sandbox: moving my open-source agent to Hermes on OpenShell](https://heybrahim.com/notes/openshell-sandbox-inside-a-sandbox.html) (2026-10-24): Inside one OpenShell sandbox, a process the agent starts gets the agent's network rules. So when I started moving Forge, my open-source agent, to Hermes on OpenShell, its workers needed a second sandbox of their own. Markdown: https://heybrahim.com/notes/openshell-sandbox-inside-a-sandbox.md - [Tokenomics: don't count tokens, count what they bought](https://heybrahim.com/notes/tokenomics-dont-count-tokens.html) (2026-10-24): J.R. Storment of the new Tokenomics Foundation says a token count is the wrong way to measure AI. Measure the cost of each outcome, count the whole bill under the models, and show the value in words the CFO already uses. Markdown: https://heybrahim.com/notes/tokenomics-dont-count-tokens.md - [Hope isn't security: the five parts of an enterprise agent](https://heybrahim.com/notes/hope-isnt-security-enterprise-agents.html) (2026-10-23): At GTC Berlin, Jensen Huang split an AI agent into five parts: the model, the harness, the sandbox, your company's knowledge and your tools. For companies getting ready for agents, he starts with the sandbox, because today the only thing stopping an agent is a prompt asking it nicely. Markdown: https://heybrahim.com/notes/hope-isnt-security-enterprise-agents.md - [Your first users are a search problem](https://heybrahim.com/notes/first-users-are-a-search-problem.html) (2026-10-19): Your first users come from finding the few who love being early or have a burning problem, not from convincing a crowd. Notes on Y Combinator's How To Get Your First Users, where Ankit Gupta replaces the minimum viable product with a minimum evolvable one. Markdown: https://heybrahim.com/notes/first-users-are-a-search-problem.md - [A customer care agent on MuleSoft Agent Fabric, where the model only talks](https://heybrahim.com/notes/mulesoft-agent-fabric-care-desk.html) (2026-10-18): Where's my sofa? is a care desk agent on MuleSoft Agent Fabric. The model only reads each message and writes each reply, while fixed steps and two Mule apps decide whose order it touches. Markdown: https://heybrahim.com/notes/mulesoft-agent-fabric-care-desk.md - [OpenClaw and Hermes Agent ship the same features with different defaults](https://heybrahim.com/notes/hermes-agent-vs-openclaw.html) (2026-10-16): OpenClaw and Hermes Agent are open-source AI agents that live on your own machine and answer you in your chat apps. Their features now match, but OpenClaw is set up for reach, Hermes for learning, and both run tools on your computer by default. Markdown: https://heybrahim.com/notes/hermes-agent-vs-openclaw.md - [A preceptor's job is to ship seniors](https://heybrahim.com/notes/preceptors-ship-seniors.html) (2026-10-16): In PRAISE, a Microsoft and Carnegie Mellon program, the senior engineer gets a label from nursing: preceptor, someone whose job is to make new seniors. Both expect AI to stop companies hiring juniors, and this is their answer, from Scott Hanselman's conversation with Michael Hilton and Julia Gao. Markdown: https://heybrahim.com/notes/preceptors-ship-seniors.md - [Review the plan, not the pull request](https://heybrahim.com/notes/review-the-plan-not-the-pull-request.html) (2026-10-15): Overmind stopped reading its own pull requests and now reviews a plan of about a page and a half, before an agent writes any code. Dylan Ratcliffe's talk on why, and how lead time fell from 3.2 days to 1.3. Markdown: https://heybrahim.com/notes/review-the-plan-not-the-pull-request.md - [One command to make GitHub Copilot read your Jira, and see what every answer costs](https://heybrahim.com/notes/pilotkit-copilot-jira-confluence.html) (2026-10-12): pilotkit connects a GitHub Copilot license to Jira and Confluence with one command. Copilot gets read-only tools, the API token stays in the OS keychain, and a local chat page shows the AI credits each answer costs. Markdown: https://heybrahim.com/notes/pilotkit-copilot-jira-confluence.md - [Fork your agent's harness, keep the winner](https://heybrahim.com/notes/fork-your-agents-harness.html) (2026-10-11): When a coding agent stops improving, have an agent write several versions of its harness, run them on the same task, and keep the best. Fausto Albers showed how at AGNTCon + MCPCon Europe, with the eval outside the loop and the cost per accepted change in view. Markdown: https://heybrahim.com/notes/fork-your-agents-harness.md - [Tailscale Serve: HTTPS for 50 self-hosted apps, none of them on the internet](https://heybrahim.com/notes/tailscale-serve-self-hosted-apps.html) (2026-10-09): Every app on my Mac mini listens on 127.0.0.1 only, and tailscale serve gives each one an HTTPS URL that only my own devices can reach. There are no open ports, no reverse proxy to maintain, and the phone gets real certificates. Markdown: https://heybrahim.com/notes/tailscale-serve-self-hosted-apps.md - [Put your agents where your team can reach them](https://heybrahim.com/notes/dont-be-your-teams-meat-proxy.html) (2026-10-08): At OpenAI DevDay, OpenClaw's Peter Steinberger showed that when your agents run on your own machines, teammates can only reach them through you. His fix is a shared team server where anyone can read and prompt anyone's agent session. Markdown: https://heybrahim.com/notes/dont-be-your-teams-meat-proxy.md - [TokenOps: why a 40-turn agent costs 16×, not 4×](https://heybrahim.com/notes/tokenops-agent-cost-curve.html) (2026-10-08): An agent re-sends its whole context on every turn, so at 40 turns its uncached input costs 16 times what 10 turns cost, not 4 times. Gareth Bland's TokenOps talk at Microsoft explains that curve, a three-ratio score for what tokens bought, and when to stop spending. Markdown: https://heybrahim.com/notes/tokenops-agent-cost-curve.md - [AI coworkers in Slack need their own identity](https://heybrahim.com/notes/ai-coworkers-slack-agent-identity.html) (2026-10-07): An AI agent that works in Slack like a colleague needs its own identity, not a human's login. Jiquan Ngiam shows how MintMCP runs 40 to 50 such agents: shared threads, permissions decided per message, memory split by audience. Markdown: https://heybrahim.com/notes/ai-coworkers-slack-agent-identity.md - [Websites won't die, they'll stop being the front door](https://heybrahim.com/notes/websites-lose-the-front-door.html) (2026-09-27): Liad Yosef, co-creator of MCP Apps, says personal assistants will replace websites as the main way into the web. I checked my own site against his research: my MCP server sat two clicks deep, behind the one file agents skip. Markdown: https://heybrahim.com/notes/websites-lose-the-front-door.md - [What breaks when an AI agent runs a real business](https://heybrahim.com/notes/what-breaks-when-an-agent-runs-a-business.html) (2026-09-27): Prosus let an AI agent run six real vending machines in Amsterdam for four months, and the business lost money. It tested on live machines, bought stock that didn't fit and cut prices until the profit was gone. Markdown: https://heybrahim.com/notes/what-breaks-when-an-agent-runs-a-business.md - [The agentic software factory is the 2022 one with one box swapped](https://heybrahim.com/notes/software-factory-2022-to-lights-off.html) (2026-09-24): Coding agents replaced the builder in the software factory that teams already ran in 2022, and kept every other box. Building now takes minutes instead of days, so the slow part of shipping moved to review. Markdown: https://heybrahim.com/notes/software-factory-2022-to-lights-off.md - [Counting Claude Code tokens from the logs? You're probably 2.5× too high](https://heybrahim.com/notes/claude-code-token-overcount.html) (2026-09-18): Claude Code logs each part of a reply on its own line, and every line repeats the whole reply's token usage. Counting lines overcounts tokens about 2.5× and output about 4.4×, so key each reply by its message id, as Meterlex does. Markdown: https://heybrahim.com/notes/claude-code-token-overcount.md - [The official MuleSoft MCP server can't read your logs, so I built one that can](https://heybrahim.com/notes/mulewatch-anypoint-logs-mcp.html) (2026-09-18): mulewatch is an open-source MCP server that lets Claude, Cursor or any MCP client read your MuleSoft Anypoint logs, live and archived, without writing anything. Its main trick is finding the right replica in the Monitoring Archive in one call instead of hundreds. Markdown: https://heybrahim.com/notes/mulewatch-anypoint-logs-mcp.md - [My README generator isn't allowed to make things up](https://heybrahim.com/notes/vibetidy-grounded-readme.html) (2026-09-18): vibetidy writes READMEs for AI-generated repos, but the LLM only turns facts that a scanner verified into prose. The prompt forbids it to invent a feature, a flag, an environment variable, a dependency or a version. Markdown: https://heybrahim.com/notes/vibetidy-grounded-readme.md ### Short notes - [Five hallway claims about personal agents, checked](https://heybrahim.com/notes/five-hallway-claims-about-personal-agents.html) (2026-10-25): Sid, a builder on the AGNTCon + MCPCon NA livestream, made five claims about personal agents. Three hold up as he said them, and two go too far: the harness doesn't beat the model, and failures don't compound "exponentially". Markdown: https://heybrahim.com/notes/five-hallway-claims-about-personal-agents.md - [macOS lied about my Docker memory usage](https://heybrahim.com/notes/mac-memory-footprint-not-rss.html) (2026-09-26): macOS denies USS for almost every process, so my Mac mini dashboard was sizing them by RSS, which overstates a VM's real cost. The fix is physical footprint, the number Activity Monitor shows. Markdown: https://heybrahim.com/notes/mac-memory-footprint-not-rss.md - [Why my transcription service ballooned to 42 GB](https://heybrahim.com/notes/mlx-buffer-cache-leak.html) (2026-09-26): MLX's buffer cache has no cap by default and never clears itself, so my long-running Whisper service kept growing. A cache limit, a cleanup after each job and an idle unload fixed it. Markdown: https://heybrahim.com/notes/mlx-buffer-cache-leak.md - [I put my portfolio behind an MCP server](https://heybrahim.com/notes/portfolio-mcp-server.html) (2026-09-26): api.heybrahim.com/mcp is a stateless MCP endpoint, so any AI agent can look me up the way a recruiter would. Five tools read the live site, and a sixth sends a CV request. Markdown: https://heybrahim.com/notes/portfolio-mcp-server.md ## Contact - Email: b.bousnguar@gmail.com - LinkedIn: https://www.linkedin.com/in/brahim-bousnguar/ - GitHub: https://github.com/brbousnguar - Credly (SAP certification badges): https://www.credly.com/users/brahim-bousnguar