~/arnArnaud Aldebert
agent onlineLet’s talkFR

~/portfolio · agent runtime

ArnaudAldebert

Fullstack & AI Systems Developer

From front to cloud, all the way to agents: tested, deployed, maintainable applications.

Nîmes, France · remote friendly
01 · retrieve()About

Context retrieved

What RAG surfaces about the candidate

Good code reads like a good prompt: clear, typed and unambiguous. I build systems that stay readable — for humans, tests and agents.

Bac+5▸ ESIEA Master’s degree
3+▸ years of experience
Fullstack▸ front · back · DevOps
FR/EN▸ professional bilingual

Fullstack developer passionate about software architecture and development best practices. My expertise focuses on building robust applications and keeping them maintainable for the long run.

In charge of technical standardization across critical environments, I favor tested, high-performance code to turn technical complexity into operational stability.

Retrieved documents

Identity
Arnaud Aldebert — Nîmes, France
Role
Fullstack & AI Systems developer · front, back, DevOps
Experience
3+ years — banking, services & real estate
Education
ESIEA Master’s (MSc) — RNCP level 7
Languages
FR · EN (TOEIC 910)
Based in
Nîmes, France · remote friendly
02 · tools()Services

Exposed tools

What I can build for you — four tools, four proofs
search_private_data()

Search & assistant over private data

The problem
Thousands of unusable internal documents, and a ban on sending them to a third-party service.
What I deliver
Ingestion, vector indexing, hybrid semantic and lexical search, sourced answers with verifiable citations, strict per-user isolation — plus retrieval quality measured continuously, so a regression shows up before release.
Proof
KeepForLater — pgvector/HNSW fused via RRF with Postgres full-text, citations highlighted at the exact passage, search that crosses languages, replayable evaluation harness in CI.
RAGpgvectorEmbeddingsHybrid searchEvals
run_llm_workflow()

Reliable, measured LLM workflows in production

The problem
AI jobs that crash halfway, emit invalid JSON, cost a fortune to re-run — and that nobody can price precisely or tell whether their quality has drifted.
What I deliver
Durable pipelines resuming at the failed step, schema-validated structured outputs, tools exposed over MCP. Plus the apparatus around it: cost, latency and prompt version traced on every call, versioned prompts with a drift guard, automated evaluations as a regression gate.
Proof
KeepForLater — five checkpointed steps, Zod end to end, MCP server, one telemetry row per call emitted (failures included), traces exported to Langfuse, separate staging and production environments.
LLM workflowsStructured OutputsMCPLLMOpsAI FinOps
harden_app()

Application modernization & hardening

The problem
A critical application nobody dares to touch: untested, slow, expensive to evolve.
What I deliver
Debt audit, strict typing, dead-code removal, full test strategy, load-time optimization, automated deployments. The same standard applies to AI components: what is not measured cannot be maintained.
Proof
La Poste IT — from 0% to over 90% coverage, 6000+ tests, load time halved on critical applications.
Audit & refactoringTest strategyCI/CDAngular
ship_product()

End-to-end fullstack application

The problem
A product idea, no engineering team, and the need for someone to own it all the way to production.
What I deliver
Architecture, front end, API, database, hosting, CI/CD, go-live and knowledge transfer.
Proof
Mazarin International Real Estate — full solution shipped solo, from architecture to deployment; KeepForLater — live product with billing and GDPR compliance.
AngularReactSpring BootServerless / Edge
Discuss your needs ↗
03 · analyze()Skills

Stack analysis

Seven modules loaded in memory

import Web

HTMLCSSTailwindJavaScriptTypeScriptAngularReactViteJavaSpring BootNode.jsBunHono.jsHono RPCZodBetter AuthStripe

import Mobile

JavaKotlinFlutterIonic / Capacitor

import DevOps / Cloud

CI/CDGitDockerKubernetesGitLab CIArgo CDGitHub ActionsJenkinsScalewayCloudflare Workers / WorkflowsServerless / EdgePostHogSentry

import Databases

SQLMySQLPostgreSQLNeonpgvectorDrizzle ORMPrisma ORMNoSQLMongoDB

import AI & GenAI

LLMRAGEmbeddingsSemantic searchChunkingPrompt engineeringStructured OutputsLLM workflowsAgentsMCPVercel AI SDKGeminiMultimodalTTSLLMOpsLangfuseLLM observability, evaluation & calibrationAI FinOps

import AI DevTools

Claude CodeCursorCopilotCodexMCPSkills

import Project management

JiraTrelloAgileScrumKanban
04 · build()Experience & projects

Production deployments

Solo-built projects and client missions shipped — build logs included

projectsWhat I build

keepforlater/app:prod#412 · passedIn production

KeepForLater — AI second brain

Personal product · design & development

SaaS platform turning information overload into queryable knowledge: it ingests long-form sources (YouTube, podcasts, articles, threads, newsletters), distills them with AI, indexes them as vectors and lets you chat with your own knowledge base. Designed, built and shipped solo end to end — architecture, product, edge, AI, billing and compliance.

Metrics
  • Durable LLM workflow (Cloudflare Workflows): extract → distill → categorize → embed → finalize, each step checkpointed, idempotent and retried independently — if embedding fails, transcription is not re-billed.
  • Strict Structured Outputs: the whole distillation (TL;DR, key points, concepts, narrative summary, timestamped chapters) is emitted as Zod-validated JSON via the AI SDK, with a self-assessed groundedness score surfaced in the UI.
  • Hybrid RAG: dense search (pgvector, HNSW index, 512-d embeddings) fused via RRF with Postgres full-text lexical search, overlapping semantic chunking, per-source diversification, SSE-streamed chat with verifiable citations and exact passage highlighting.
  • Cross-lingual FR ⇆ EN semantic search, measured: replayable RAG eval harness in CI.
  • Native multimodal transcription: long media are segmented via byte-range requests then transcribed by Gemini, working around the model output limit.
  • MCP server (Streamable HTTP): the second brain exposed as tools — search_sources, get_source, save_url — usable straight from Claude, ChatGPT or Cursor.
  • Active retention: on-demand quizzes (MCQ, true/false, flashcards) as structured output, SM-2 spaced repetition with streaks and weak topics, Obsidian-like knowledge graph (co-occurrence and semantic similarity), 30-voice TTS with edge cache and a private podcast feed.
  • Monitoring automations: 7 watch types (YouTube channels and playlists, subreddits, RSS feeds, keywords, weekly digest) run by Workers cron with transactional emails.
  • Three product surfaces: SSR web + installable PWA (light offline, Web Push), cross-browser WXT extension (Chrome MV3, Firefox) and MCP server — Bun monorepo fully typed end to end via Hono RPC and shared Zod schemas.
  • Controlled AI economics: per-action costs measured via usageMetadata, credit system calibrated on real COGS, Stripe subscriptions and credit packs, soft paywall; PostHog/Sentry loaded only after GDPR consent.
  • Full LLMOps: one telemetry row per call emitted — failures included, since they are billed — a versioned prompt registry with a CI drift guard, free deterministic evaluations on every contribution plus an independent LLM judge nightly, traces exported to Langfuse from the edge runtime, separate development, staging and production environments.
Stack
Bun WorkspaceReact 19Hono RPCCloudflare WorkersWorkflowsAI SDKGeminiHybrid RAGpgvectorNeon PostgresDrizzleZodMCPBetter AuthStripeWXTLLMOpsLangfuse
Read the case study →Inspect the deployment ↗
mcp-rncp/worker:prod#58 · passedIn production

mcp-rncp — the RNCP inside AI assistants

Open-source project · design & development

MCP (Model Context Protocol) server bringing the French national registers of professional certifications (RNCP, RS — France compétences open data) straight into Claude, ChatGPT or Cursor: sheet validity, EDOF skill blocs, SIRET habilitation, search by level or field. Free, no API key, every answer sourced and dated. Designed, built and hosted solo: ingestion, search engine, npm CLI, Cloudflare Worker, website.

Metrics
  • Automated weekly ingestion: 570 MB of XML parsed as a stream, a 30,000-sheet SQLite/FTS5 database rebuilt in under a minute, published as a GitHub release and loaded into Cloudflare D1 with zero downtime (blue/green databases).
  • One tool codebase (MCP SDK v2, Zod) for two runtimes: npm stdio CLI (node:sqlite) and a Cloudflare Worker over Streamable HTTP with rate limiting, caching and JSON logs.
  • BM25 full-text search with French synonyms evaluated in CI on 46 real questions: 96% recall@5, 100% exactness on validity/habilitation checks.
  • Published on npm and the official MCP registry; SEO/GEO website (structured data, llms.txt) server-rendered without JavaScript.
Stack
MCPTypeScriptBun WorkspaceCloudflare WorkersD1SQLite FTS5HonoZodGitHub ActionsnpmCI evals
Inspect the deployment ↗

experienceProfessional experience

laposte/front:prod#2647 · passedIn production

Front-End standardization — La Poste

Angular Front-End Developer · Klanik

Angular front-end developer mission within La Poste IT, focused on standardizing the ecosystem across several repositories (Offers, Contracts, Products). I drove technical decisions on my scope to cut debt, harden critical applications and establish durable standards.

Metrics
  • Audit, refactoring and clean code: strict typing, dead-code removal, Angular 19 best practices.
  • Modernized reactivity with Signals and reworked loading patterns (−50% page load time).
  • Full QA strategy: Jasmine/Karma setup, from 0% to 90%+ coverage (+6000 tests).
  • Automated deployments via Jenkins, AI-assisted best practices and Scrum teamwork.
Stack
Angular 19SignalsJasmine / KarmaJenkins
mazarin/fullstack:prod#312 · passedShipped

Mazarin International Real Estate

Freelance mission · Sole proprietorship

Designed and deployed a full-stack solution for a prestige real-estate agency in Paris. From architecture to production, I owned the entire stack to deliver a reliable, high-performance product ready for day-to-day business use.

Metrics
  • High-performance responsive showcase site in Angular 20, focused on conversion and a premium brand image.
  • Java 21 / Spring Boot 3.5+ API for transactional email handling and business logic.
  • Serverless container architecture on Scaleway, domain management and Google Workspace setup.
  • GitHub Actions CI/CD pipelines for automated build and production deployment.
Stack
Angular 20Java 21Spring BootScalewayGitHub Actions
ca/portail:prod#1893 · passedShipped

Client portal — Crédit Agricole

Web developer · CATS Montpellier

Two years of missions at Crédit Agricole Technologies & Services on critical banking environments. I contributed to the public website, new customer journey integration and the microservices portal, ensuring performance and production reliability.

Metrics
  • Enhancements to the customer-facing website for retail and business clients.
  • UI for integrating new digital customer journeys (accounts, loans, savings, insurance).
  • Microservice APIs for the new portal: authentication, simulators, contract management.
  • Diagnosis and fixes on test and production environments, working in Agile/Scrum.
Stack
Angular 17–19Java 11–21Spring BootAEMKubernetesArgo CD
05 · pipeline()How I work

The pipeline

From the documentation — four stages, zero exceptions
01

Understand the business

Read the need before writing a line. The best code is the code never written — the second best solves the real problem.

02

Test everything

Strict typing, automated tests, systematic reviews: trust is built in the pipeline, not in production. Untested code is debt.

03

Automate everything

From continuous integration to deployment: every commit goes through the pipeline. Humans review, machines do the routine.

04

Maintain for the long run

Maintainability beats the quick win. Code still readable in five years — by a human or an agent.

06 · fine_tune()Learning — tech, concepts & certifications

Fine-tuning in progress

The model keeps training
React
Actively training
Artificial intelligenceGenAI, LLM, RAG, Agents, Workflows
Actively training
UnityGame development — Udemy course
Queued
BlenderUdemy course
Queued
Three.js / React Three Fiber
Queued
Certified Mid-Level Angular DeveloperCertificates.dev
Queued
07 · cron()Passions & interests

Background jobs

What runs outside office hours

Computing

Development, AI, Cybersecurity

next run in --:--

Investing

Crypto, Stock market

next run in --:--

Video games

Fan of teamon Counter-Strike

next run in --:--

Japanese culture

Language, traditions and aesthetics

next run in --:--

Cinematography

Art direction, visual storytelling and composition

next run in --:--
08 · deploy()Contact

Deploy Arnaud

A developer ready for your production
Environment
Team, product or mission — we shape the setup together
Image
arnaud:latest — fullstack + AI
Replicas
1 — single instance
SLA
Reply within 24h
Discuss your project ↗Download the resume (PDF)
ESIEA
Master’s degree (MSc) — Software engineering expert (RNCP 35419 — Level 7)
Certifications
Certified Junior Angular Developer — Certificates.dev · Scalable Microservices for Developers — Vanderbilt University · TOEIC: 910 pts · English4IT · Voltaire certificate: 613 pts

Or the channel of your choice

LinkedIn↗Professional DMs welcome. Track record, recos, network.GitHub↗Code in the open. Maybe this very site.Malt↗Freelance missions — verified profile.Email↗The most direct channel. Reply within 24h.arnaudtishu@gmail.com