Open to AI architecture roles

Hosam MahmoudAI Systems Architect

I design and ship agentic AI systems end to end: multi-agent pipelines, retrieval-augmented generation, and self-hosted inference infrastructure. My open-source systems run entirely on local models, with no external API dependency and no per-call vendor cost.

Currently architecting LLM-powered automation at Capgemini. M.Eng. in AI & Data Science, University of Ottawa.

Book a 30-min call See the systems GitHub
Cairo, Egypt AWS Certified ML Specialty Arabic & English
Hosam Mahmoud Ibrahim
Flagship systems

Systems I built and open-sourced

Not demos or notebooks. Complete systems with architecture, safety rails, operations tooling and tests. Both run their inference locally.

Agentic pipeline · Python

Growth Engine: nine-agent outbound automation

An autonomous B2B prospecting system. Nine specialised agents discover leads, research them, write personalised outreach with a local LLM, send it under hard-coded deliverability rules, then watch for replies and report on the funnel.

Every stage is an independently testable unit that passes state through a shared persistence layer, so any single agent can be swapped, re-run or reasoned about in isolation.

FastAPIAPSchedulerSQLAlchemy 2.0AlembicOllamaaiogram 3PostgreSQLPytest
View repository →
Growth Engine architecture A provider-agnostic LLM layer sits above nine sequential agents (prospector, enricher, verifier, researcher, writer, sender, reply watcher, reporter and dashboard) which all read and write through a shared persistence layer. LLM layer · per-role model assignment Ollama (local, default) · OpenAI · Anthropic · Groq · OpenRouter · daily caps with automatic local fallback 01PROSPECTOR license registries · maps · CSV 02ENRICHER crawls sites for contacts 03VERIFIER syntax · MX · SMTP probe 04RESEARCHER summarise · deep web research 05WRITER personalised · style-enforced 06SENDER warm-up · windows · jitter 07REPLY WATCHER IMAP poll · classify · alert 08REPORTER funnel metrics 09DASHBOARD config · monitoring · control Shared persistence · state passed between every stage SQLAlchemy 2.0 · Alembic auto-migrations · SQLite → PostgreSQL · nightly backups

Nine agents, one shared state layer, one swappable model layer.

Compliance built into the architectureWarm-up ramps, randomised send jitter, restricted send windows, suppression lists, CAN-SPAM footers, and a circuit breaker that halts a campaign above a 3% bounce rate.
Three-tier execution safetyDRY_RUN writes files without sending, SANDBOX delivers only to a test inbox, and LIVE demands two-key confirmation: one environment variable plus a Telegram gate.
Provider-agnostic by designEach role gets its own model. Local Ollama is the default, so the system runs at zero marginal cost; cloud providers are optional, capped daily, and fall back to local.
Operable, not just runnableAutomatic schema migrations, nightly backups, health-check endpoints, and an offline test suite that mocks every network and LLM call.
Multimodal pipeline · Python

YouTube Shorts Studio: self-hosted video AI pipeline

Turns one long-form video into multiple publish-ready vertical Shorts. Transcription, semantic segmentation, montage editing, face-tracked reframing, right-to-left Arabic captioning, thumbnail generation and upload, all on hardware you already own.

The interesting constraint was memory: Whisper, the LLM and the background-removal model all want the GPU at once. Scheduling them through a shared VRAM budget lets the whole stack run on a single 8 GB consumer card.

Whisper large-v3OllamaCommand-RQwen 2.5FastAPIOpenCVFFmpegrembgyt-dlp
View repository →
YouTube Shorts Studio pipeline A ten-stage pipeline running from input acquisition through transcription, semantic segmentation, montage, reframing, captioning, overlay, metadata, thumbnail and upload, with Whisper, the LLM and background removal sharing one 8 GB GPU through scheduled VRAM budgeting. Single 8 GB GPU · models scheduled through a shared VRAM budget Whisper large-v3 · local LLM via Ollama · rembg background removal, never resident simultaneously 01INPUTyt-dlp fetch 02TRANSCRIBEword timestamps 03SEGMENTLLM · sentence idx 04MONTAGEsilence removal 05REFRAMEface track 9:16 06CAPTIONRTL Arabic 07OVERLAYsubscribe anim 08METADATAtitle · description 09THUMBNAILface cutout 10UPLOADData API v3

Ten stages, one GPU, no cloud services and no per-video cost.

Indices, not timestampsThe segmentation LLM returns sentence indices rather than times. Cuts then always land on semantic boundaries, removing the dominant failure mode of timestamp-based approaches.
Right-to-left captioningWord-by-word Arabic highlighting synchronised to Whisper word timestamps, with correct bidirectional text layout and shaped glyphs.
Thumbnails from real pixelsFace detection plus background removal composites the actual presenter over three style templates, rather than generating a face that was never in the video.
Hardened for exposureHMAC cookie authentication, an AES-256-GCM vault for API keys, SSRF-guarded downloads and strict Content Security Policy headers, because it is reachable from a phone.
Multimodal assistant · Docker

Multimodal Telegram AI Assistant: fully local, zero external APIs

A Dockerised assistant handling text conversation, image captioning, image transformation and text-to-image generation entirely on local open-source models. n8n orchestrates the workflow; PostgreSQL persists conversational memory so context and personalisation survive across sessions.

Because nothing leaves the machine, it is private by construction and costs nothing per message.

LLaMAGemmaComfyUI Fluxn8nPostgreSQLDocker
Capabilities

What I work with

Depth in the generative and agentic stack, backed by the data and cloud engineering foundations that keep these systems running in production.

Generative AI & LLMs

  • CoreRAG, prompt engineering, structured output, function & tool calling, Model Context Protocol
  • TuningQLoRA, LoRA, PEFT, SFT, quantization (GGUF, Q4)
  • RetrievalEmbeddings, semantic search, vector databases, evaluation & guardrails
  • ModelsLLaMA, Gemma, Qwen, Command-R, Mistral, Whisper large-v3, OpenAI, Anthropic

Agentic systems & orchestration

  • FrameworksLangChain, LangGraph, multi-agent pipelines, agent state machines and routing
  • Workflown8n, APScheduler, event-driven architecture, human-in-the-loop approval gates
  • ResilienceProvider fallback, rate limiting, retry and circuit-breaker patterns

AI infrastructure & MLOps

  • ServingOllama, Hugging Face Transformers, ComfyUI, GPU & VRAM scheduling
  • DeliveryDocker, Docker Compose, FastAPI, REST, webhooks, Cloudflare Tunnel, CI/CD
  • Operationsdev/staging/prod promotion, Pytest, logging, monitoring, health checks

Data, cloud & analytics

  • StoragePostgreSQL, SQLite, SQLAlchemy 2.0, Alembic, Cassandra, MongoDB
  • ScaleApache Spark, Apache Kafka, Azure Synapse Analytics, ETL & ELT pipelines
  • CloudAWS, Microsoft Azure, Power BI, data visualisation
  • LanguagesPython, SQL, R, Scala, PyTorch, TensorFlow, scikit-learn, OpenCV
3+ yrsProduction AI & data engineering
6 hrs → secReporting process automated away
15Public repositories
8 GBGPU running a full multimodal stack
Experience

Where I have worked

AI & Automation Engineer

09/2025 to Present

Capgemini · Cairo, Egypt

  • Architect and deploy end-to-end AI automation systems connecting multiple business applications, replacing manual processes with orchestrated, monitored pipelines.
  • Design and build intelligent chatbots and voice assistants for client-facing platforms, owning conversational flow, state handling and interaction models.
  • Integrate LLMs into chatbot architectures to improve contextual understanding, grounding and response quality across multi-turn conversations.
  • Maintain development, staging and production environments, managing code migration and deployment for stable, consistent releases.
  • Author the technical documentation and workflow architecture diagrams that enable handover and reuse across teams.

Data Analyst

06/2023 to 09/2025

Capgemini · Cairo, Egypt

  • Automated recurring manual Excel reporting with Python, reducing a six-hour process to seconds.
  • Developed SQL database solutions powering real-time Power BI dashboards, accelerating operational decisions.
  • Built dashboards that turned raw data into actionable insight, with UI/UX refinements for a presentation-ready result.
  • Collected and preprocessed data from heterogeneous sources, then analysed it to surface patterns, trends and anomalies.

Education

M.Eng. Artificial Intelligence & Data Science

University of Ottawa, Canada

02/2022 to 02/2023 · Grade: Excellent

B.Sc. Computer Science

Misr University for Science & Technology

09/2017 to 07/2021 · GPA 3.80 / 4.00

AWS Certified Machine Learning SpecialtyAmazon Web Services · May 2022 AWS Certified Cloud PractitionerAmazon Web Services · February 2022 Microsoft Power BIOctober 2023 Dale Carnegie Soft SkillsMay 2022 · also Berlitz English, ICDL
Open source

Research & engineering archive

Thirteen repositories from my M.Eng. and undergraduate work, each pairing the source code with the reports, papers and presentations that document it.

AI-For-CybersecurityCT-GAN intrusion detection, Kafka model streaming, DCGAN anomaly detection Language-Identification-SystemSeven-language speech ID: dataset pipeline, MFCC/Mel modelling, Flask deployment Smart-Cities-SimulationCrowdsensing pipeline: task generation, CGAN synthesis, anomaly detection, ns-3 NLP-ProjectsAuthorship attribution, book clustering, ontology-driven chatbot Big-Data-SparkDistributed processing in Scala, PySpark and Spark SQL with streaming jobs Computer-VisionTransfer-learning classification plus KNN and filters built from scratch Machine-Learning-From-ScratchDecision trees, boosting, KNN and OVR/OVO from first principles Data-Science-AnalyticsSentiment-driven price prediction, churn analysis in R, SQL analytics Deep-LearningNeural architectures on the UNR-IDD intrusion detection dataset Azure-Synapse-StreamingCloud warehousing and real-time event processing on Azure Reinforcement-Learning-ProjectsAgent trained to drive through environment interaction Social-Media-ScraperBatch-downloads 100+ videos per run with automatic trimming Mongodb-CassandraDocument versus wide-column NoSQL data modelling
Contact

Let's talk about your AI systems

Whether you are scoping an agentic architecture, moving an LLM prototype into production, or want to cut inference cost by self-hosting, book a slot and let's get into it.

Emailhosam2mahmoud@gmail.com LinkedInhosam-mahmoud-ibrahim GitHubgithub.com/HosamBadawi
Book a 30-minute call
Pick any slot that works for you, with no back-and-forth email.
Open in Calendly ↗
Loading the booking calendar…
Open calendly.com/sam_mahmoud directly