Atomic facts, not chunks
Chunking by token window splits a claim from its context. Indexing extracted facts gives the retriever units that stand on their own, so a generated report traces back to a source.
Available for freelance & full-timeAI · RAG · Agentic systems
I build RAG and multi-agent systems that run in production.
I work on RAG architecture, LangGraph agent orchestration, and the GCP infrastructure under both. My graduation project runs inside OCP Group. I am currently building a hybrid-RAG platform on GCP for a client.
- Casablanca, Morocco
- ibrahimgoumrane01@gmail.com
- 1st
- Class rank at ENSAM Casablanca, certified
- OCP
- RAG platform in internal production at OCP Group
- <300ms
- p95 on a live delivery eligibility engine
- 6
- Freelance clients delivered end to end
What I do
Most of my work is retrieval design. How documents get decomposed, indexed, retrieved and reranked decides whether a model answers reliably over a corpus it never saw in training.
On top of that I build the agent orchestration in LangGraph, with error recovery and human validation checkpoints. Underneath, the GCP infrastructure that runs it: Terraform, Cloud Run, Cloud SQL, Vertex AI.
I work freelance and solo, so I also scope the engagement, price it, and renegotiate when the spec moves.
Flagship case study
OCP SolutionsLive
Groupe OCP subsidiary. Graduation project, defended 22 June 2026.
A multi-agent platform that writes documentary-analysis reports
- Role
- Agentic AI Engineer
- Period
- March – August 2026
Problem
Analysts wrote documentary-analysis reports by hand from large, mixed-format internal corpora. A naive RAG pass over that material returns plausible text with no traceable grounding, which is useless for a report someone has to sign.
Approach
LangGraph orchestration behind FastAPI, with a React supervision UI over WebSocket. Retrieval indexes extracted facts, questions and atomic facts rather than raw chunks, in Milvus with Qwen Embedding 4B, retrieved by bi-encoder and reranked with BAAI/bge-reranker-v2-m3.
Outcome
Deployed on OCP servers and used internally by the group. Supervisors: Pr. Samih Lazaiz (ENSAM), M. Mohammed Zouai (OCP).
Pipeline
- 01
Ingest
Multimodal internal documents
- 02
Decompose
Facts, questions and atomic facts, not raw chunks
- 03
Index
Milvus, Qwen Embedding 4B
- 04
Retrieve
Bi-encoder over the decomposed index
- 05
Rerank
BAAI/bge-reranker-v2-m3
- 06
Orchestrate
LangGraph agent graph with error recovery
- 07
Validate
Human checkpoints inside the graph
- 08
Stream
WebSocket updates to the React UI
Key decisions
Reranking as its own hop
The bi-encoder is fast and favours recall. The cross-encoder reranker is slow and favours precision. Two stages avoid compromising on a single similarity threshold.
Checkpoints inside the graph
Validation at the end of a pipeline only reports that a run was wrong. As nodes inside the graph, checkpoints stop a bad intermediate state before it reaches the rest of the report.
Stack
- LangGraph
- FastAPI
- Milvus
- Qwen Embedding 4B
- bge-reranker-v2-m3
- React
- WebSocket
- Python
Currently building
SeementIn development
fca-agentic-outlook
A hybrid-RAG platform on GCP, provisioned in Terraform
Cloud / AI Infrastructure Engineer, freelance
I own the GCP resource stack for an agentic RAG platform, plus part of the retrieval strategy. The engagement is active.
Stack
- Terraform
- GCP
- Vertex AI RAG Engine
- pgvector
- Cloud Run
- Cloud SQL
- GitLab CI
- Workload Identity Federation
Hybrid retrieval: Vertex AI RAG Engine + pgvector
In developmentVertex AI RAG Engine does strong semantic search over a large corpus, but its native filtering and metadata search are limited. pgvector handles the filtered queries. Combining both beats either alone.
The whole GCP stack as Terraform
In developmentCloud Run, Vertex AI RAG Engine, Cloud SQL and storage buckets, defined as code and applied from GitLab CI. The environment is reproducible instead of clicked together in a console.
CI auth without service-account keys
DeliveredGitLab CI authenticates to GCP through Workload Identity Federation instead of a long-lived JSON key. Nothing secret sits in the repository or the CI variables.
Private Cloud SQL through impersonation
DeliveredThe Auth Proxy reaches a private Cloud SQL instance by impersonating a service account, so no developer account gets direct database access.
Case studies
Selected work
Four client engagements.
Problem
A pre-launch logistics startup had to know whether a two-hour slot could be honoured before promising it. Three systems each hold part of the answer and none agree by default.
Approach
An eligibility engine answering under 300ms at p95, with atomic stock reservation and weighted hourly capacity, on an append-only event journal with periodic cross-system reconciliation. Three production APIs: Shopify Admin/GraphQL over OAuth client credentials, Odoo XML-RPC, and Onfleet REST plus webhooks with HMAC verification.
Outcome
I negotiated the two-phase proposal (36 days, €2,650) from a 15+ page spec, pricing the gaps between stated scope and actual requirements. Also set up VPS deployment, Git CI/CD and secrets management.
Stack
- Event sourcing
- Shopify GraphQL
- Odoo XML-RPC
- Onfleet
- HMAC
- VPS
Freelance engineer, 36 days in two phases
Problem
Three organisation types share one lifecycle. Expert assessors record damage, insurers validate and release the dossier, approved scrap dealers bid. Each needs a different view, and the dossier must freeze the moment an auction opens.
Approach
A four-tier role system with a strict validation workflow, and a custom auction engine: per-bidder ceilings locked at 30% above each bidder's own opening bid, minimum increments that change in the final hour, and visibility limited to whether a bidder currently leads. All enforced server-side with row-level locking, with idempotent scheduled closure for shared hosting.
Outcome
Delivered the spec, data model, API contract, design system and implementation across two releases, so the client could validate the core workflow before funding the rest. The ceiling rule and the increment rule turned out to conflict, which I found while writing the spec.
Stack
- Next.js
- Spring Boot
- PostgreSQL
- WebSocket
- Row-level locking
Freelance engineer: spec, data model, implementation
Problem
A craft brewery needed prospects across two motions, B2B distribution and events, with no sales team to research them.
Approach
A Python orchestrator runs cross-source web research, extracts name, city, website, email, phone, contact and source, deduplicates, then scores each prospect A+ to D on potential volume, seasonality and contact reliability before hand-off into two HubSpot pipelines.
Outcome
Delivered with Lemlist sequences that auto-stop on reply, plus deliverability setup: SPF/DKIM/DMARC, progressive volume ramp-up and invalid-address cleanup. Won on a job-platform bid under €500.
Stack
- Python
- HubSpot
- Lemlist
- Lead scoring
- SPF/DKIM/DMARC
Freelance, solo: scoping through hand-off
Problem
Companies needed to study metrics by instrument, axis and theme, benchmark against their sector, and get a report an expert would sign, starting from balance sheets that arrive as documents in any currency.
Approach
Three hard parts: a formula interpreter so each instrument carries its own computation instead of being hard-coded, currency detection and conversion to one unified currency, and value extraction from balance-sheet documents. On top, generated Word reports with computed KPIs.
Outcome
I led the build as Technical Lead, client-facing, coordinating the dev team.
Stack
- Next.js
- Laravel
- Python
- OCR
- MySQL
- Nginx
Technical Lead, freelance
Timeline
Experience
Some roles run concurrently: freelance work overlaps with employment.
Cloud / AI Infrastructure Engineer
Seement2026, ongoingFreelance
Provisioning the GCP stack for an agentic RAG platform in Terraform: Cloud Run, Vertex AI RAG Engine, Cloud SQL, storage buckets, authenticated through Workload Identity Federation and wired into GitLab CI. I also own part of the retrieval strategy, a hybrid design pairing Vertex AI RAG Engine with pgvector for filtered queries.
Stack
- Terraform
- GCP
- Vertex AI
- pgvector
- GitLab CI
- WIF
Agentic AI Engineer
OCP SolutionsMarch – August 2026Graduation project
Built a multi-agent platform that automates documentary-analysis report generation over a multimodal RAG architecture, deployed on OCP servers and used internally by the group. LangGraph orchestration behind FastAPI, React supervision UI over WebSocket, retrieval on atomic-fact decomposition indexed in Milvus.
Stack
- LangGraph
- FastAPI
- Milvus
- React
- WebSocket
Prospecting automation engineer
Pierron OlivierJuly 2026Freelance, solo
Built a prospecting system for a craft brewery across HubSpot and Lemlist, on two pipelines. A Python orchestrator runs cross-source web research, deduplicates, and scores prospects A+ to D before hand-off. Also set up Lemlist sequences and email deliverability.
Stack
- Python
- HubSpot
- Lemlist
- Lead scoring
Freelance engineer
DITOCO2026, 36 daysFreelance
Built a two-hour delivery test bench for a pre-launch logistics startup, orchestrating Shopify, Odoo and Onfleet around an eligibility engine under 300ms at p95 and an event-sourced journal. Negotiated the two-phase proposal from a 15+ page spec.
Stack
- Event sourcing
- Shopify
- Odoo
- Onfleet
- VPS
Freelance engineer
Maroc Réforme2026Freelance, in development
Rebuilding a B2B salvage-vehicle auction platform for Moroccan insurers, connecting expert assessors, insurers and scrap dealers. Four-tier roles, strict validation workflow, dossier frozen once an auction opens, and a custom auction engine enforced server-side with row-level locking.
Stack
- Next.js
- Spring Boot
- PostgreSQL
- WebSocket
Automation Engineer
Top NegoceFebruary 2026Internship
Automated invoice extraction and analysis on a locally-deployed n8n instance, feeding a dashboard that supports board decisions and flags entities engaged in fraud.
Stack
- n8n
- Python
- Dashboarding
Automation & Systems Engineer
Formafast ConsultingSeptember 2025 – June 2026Part-time
The broadest role here: software, automation, data engineering and infrastructure. Refactored and deployed 10+ client sites, configured Postfix and SMTP relay for an internal mailer, built FormaFastScraper for contact enrichment, and a daily marketplace scraping pipeline on Spark. Also Linux administration, SSL/TLS and incident response, including a ransomware incident on a test database.
Stack
- FastAPI
- Apache Spark
- Selenium
- spaCy
- Linux
- Postfix
Technical Lead
FINANCIANAJune – September 2025Freelance
Financial analysis platform for studying financial health by instrument, axis and theme, with sector benchmarking and generated Word reports. The hard parts were a formula interpreter per instrument, currency detection and unification, and value extraction from balance sheets.
Stack
- Next.js
- Laravel
- Python
- OCR
- MySQL
Technical Lead
EHC → RecruiteJune – November 2025Internship → freelance
Full-stack recruitment platform covering applications, assessments and interviews. I led the technical vision, coordinated the dev team and bridged client and team. Role-based access, real-time analytics, CV building and interview prep. Deployed on a VPS with Nginx and Certbot.
Stack
- React TS
- Spring Boot
- MySQL
- Quartz
- WebSockets
Software Development Intern
Numeric WayJune – August 2024Internship
Built KanBanPro, a Jira-style project management platform with Kanban boards, real-time collaboration and chat over WebSockets, and multi-user management per project.
Stack
- Express.js
- React
- WebSockets
- Material UI
Archive
Project archive
Personal, academic and side work.
FastDev
10/2025 – 12/2025Generates, tests and deploys full-stack projects from a natural-language description, using Gemini and the Model Context Protocol. Two-stage test/production pipeline with manual validation.
- Gemini API
- MCP
- Spring Boot
- Java 21
- React
- Keycloak
- Docker
- Loki
AutoGradeAI
Automated exam grading. Tesseract pulls text off scanned papers, an LLM generates the model answer, and sentence-transformer similarity scores the student. Presented at ENSAM Casablanca's AGIOS day.
- Tesseract OCR
- OpenAI API
- Sentence Transformers
- Python
Recruite, recruitment platform
06/2025 – 11/2025Recruitment workflow from application to interview: automated assessments, role-based access for candidate, recruiter and admin, real-time KPIs, CV building and interview prep.
- React TS
- Spring Boot
- MySQL
- WebSockets
- Quartz
- JWT
- Nginx
Financiana, financial analysis platform
06/2025 – 09/2025Per-instrument formula interpretation, currency unification, balance-sheet extraction and generated Word reports. Full case study above.
- Next.js
- Laravel
- Python
- OCR
- MySQL
ArtiShare, article sharing platform
A Medium-style publishing platform built in two weeks with a classmate. Next.js front end, Laravel back end for auth, REST APIs and article management.
- Next.js
- TypeScript
- Laravel
- MySQL
RAG-Engine
07/2025 – 08/2025Builds a knowledge base from a whole website: recursive multi-page scraping with JS rendering, hybrid semantic and TF-IDF ranking, FAISS search, and an MCP server so an agent can query it.
- Python
- OpenAI API
- FAISS
- Sentence Transformers
- Playwright
- MCP
OCR model for long text sequences
11/2024 – 02/2025A CNN-LSTM model on a ResNet50 backbone, built around the CTC-loss alignment problem on long sequences. Notebook published on Kaggle.
- PyTorch
- CNN-LSTM
- CTC Loss
- ResNet50
FaceAttendance.AI
03/2025 – 06/2025Classroom attendance by facial recognition, with Redis caching so it stays usable when a whole class arrives at once.
- Django
- face_recognition
- Redis
- Next.js
- Docker
Microservices e-commerce platform
Auth, order, product, payment and file services, each with its own PostgreSQL, behind a JWT API gateway. Event-driven over RabbitMQ and Kafka, on Kubernetes after migrating off Docker Compose.
- Kubernetes
- Kafka
- RabbitMQ
- PostgreSQL
- Prometheus
- Grafana
- ELK
Rehabilitation management app
A CMS for rehabilitation agencies. Patients manage their own exercise routines, and an initial assessment tailors the programme by condition and age. Built after my own scoliosis rehab.
- React
- TypeScript
- Express
- Prisma
- MySQL
Web sales analytics pipeline
A daily Airflow pipeline that scrapes 1,000+ products, cleans and validates them, and loads PostgreSQL, with a Streamlit dashboard for price and discount tracking. CI/CD on GitHub Actions with pytest.
- Apache Airflow
- BeautifulSoup
- PostgreSQL
- Streamlit
- Docker
- pytest
FormaFastScraper, contact enrichment
Resolves full identity and contact data from a name alone. Anti-bot work throughout: Selenium migrated to nodriver, human-like delays, automatic pause on captcha. spaCy NER for extraction, geocoding for location.
- FastAPI
- React
- nodriver
- spaCy
- MySQL
- Geocoding
Marketplace scraping pipeline on Spark
A daily Apache Spark pipeline that scrapes marketplace listings, then cleans and reorganises the data before loading a remote MySQL instance.
- Apache Spark
- Python
- MySQL
Invoice extraction and fraud flagging
A locally-deployed n8n instance that automates invoice extraction and analysis, feeding a dashboard that supports board decisions and flags entities engaged in fraud.
- n8n
- Python
- Dashboarding
Toolset
Stack
Tools with real project work behind them.
Agentic & LLM
- LangGraph
- LangChain
- OpenAI API
- Google Gemini API
- Model Context Protocol
- Prompt engineering
Retrieval & vector
- Hybrid RAG design
- Vertex AI RAG Engine
- Milvus
- pgvector
- FAISS
- Qwen Embedding 4B
- bge-reranker-v2-m3
- Sentence Transformers
Cloud & infrastructure
- Terraform
- GCP (Cloud Run, Cloud SQL, GCS)
- Workload Identity Federation
- Docker
- Kubernetes
- GitLab CI
- GitHub Actions
- Linux
- Nginx
- Keycloak
Backend & APIs
- Python
- FastAPI
- Java
- Spring Boot
- Spring Security / JWT
- Node.js / Express
- Django
- Laravel
- PostgreSQL
- MySQL
- MongoDB
- WebSockets
- Event sourcing
Data engineering
- Apache Airflow
- Apache Spark
- n8n
- Selenium / nodriver
- Playwright
- BeautifulSoup
- spaCy NER
- ETL design
Observability & quality
- Grafana
- Prometheus
- Loki
- Promtail
- ELK
- pytest
Frontend
- React
- Next.js
- TypeScript
- Tailwind CSS
- Angular
Background
Education & recognition
ENSAM Casablanca
École Nationale Supérieure d'Arts et Métiers
Engineering degree, Artificial Intelligence & Software Engineering
2021 – 2026
Ranked 1st in the graduating class. Graduation project defended 22 June 2026 at OCP Solutions.
DataCamp
Data Science & Data Analyst career tracks
Certified coursework: regression, classification, statistical inference
Certification
Listed as coursework, not applied production ML.
Recognition
- Ranked 1st of the graduating class at ENSAM Casablanca, certified
- Graduation project deployed a RAG system now used internally by OCP Group
- Presented AutoGradeAI at ENSAM Casablanca's AGIOS day
- Technical Lead on several freelance engagements, client-facing