Ali Salloum

AI-Accelerated Full-Stack Architect — scalability, security, and production-readiness.

Blueprint diagram: reverse proxy to Manager, then routing to Unix sockets for supervisor and agents
25 → 1
Services to Manage
< 5ms
Proxy Overhead
CI/CD • Case Study 01

Unified Manager for Multi-Agent Deployment

I redesigned deployment into one manager per environment and automated creating agents from the admin UI—no manual systemd units, Caddy edits, or port assignment. The same model makes backups and bringing the stack up on additional servers far simpler: one service, one config, repeatable deploy.

  • Automated agent provisioning (admin UI → reload, no per-agent plumbing)
  • Easier backup and multi-server rollout (single unit + shared config)
Backend • Case Study 02

Push notifications people can rely on

I built the backend that gets notifications from your product to users’ phones, using Firebase. People register their devices in a safe, signed-in flow, and older saved tokens were moved over without breaking existing apps. When a message must reach many specific users, the system doesn’t fire everything at once—it uses a queue and a background worker so delivery stays controlled and observable.

Technical focus

  • A single background worker consumes the send queue: it processes one device at a time, updates status after each attempt, and avoids running duplicate workers in parallel.
  • Each send is recorded in an operations log; the mailing only completes when every row has a final outcome (delivered, failed, or acknowledged)—so you always know what finished and what didn’t.
Web Scraping / Automation • Case Study 03

LuukAI Scraper

Crawls a domain and turns pages into structured product and article data—one pipeline for storefronts, blogs, and mixed sites instead of a bespoke scraper per layout. Phase one walks the site, extracts detail pages first, and defers listing grids so real product pages get priority; phase two goes back and mines those deferred listings. Saves output incrementally so a run can stop and resume without starting over.

Differentiator
Reads the page—not a bespoke selector map per site
Maximum structured data per inference dollar
97%
Audited field accuracy
LuukAI promotional popup with branded header, headline, and action buttons
Live widget demo
Frontend • Case Study 04

LuukAI — Embeddable AI Assistant & Partner Widget

Partners drop in a conversational assistant: plain-language Q&A, smart recommendations, optional promos, and guided flows when needed. One codebase ships as a full web app and a floating embed—on-brand per tenant, multilingual, and built so the host site’s styling can’t break the experience.

Built with
React · TypeScript · Vite — one pipeline, multiple surfaces (app, widget, embedded shell)

Highlights

  • The widget is visually self-contained—partner pages don’t accidentally restyle or break the assistant.
  • Surveys, partner add-ons, and analytics hook in cleanly so journeys feel native, not bolted on.
  • Per-brand copy, languages (including RTL), and tailored result screens—down to industry flows and booking handoff to the host.
REACT 18MUI V5VITE 5
Scholar impact

12citations on Google Scholar

Paper (ScienceDirect)Citing articles (Google Scholar)Recent citation: npj Computational Materials (Nature Portfolio) — materials discovery using quantum-annealed ML
Publication • Case Study 05

Quantum annealing in machine learning: Qboost on D-Wave quantum annealer

Peer-reviewed paper on quantum annealing and machine learning: a QBoost-style ensemble mapped to a quadratic unconstrained binary optimization (QUBO) formulation and run on a D-Wave quantum annealer. It explores when quantum hardware can complement classical ML pipelines, how experiments are set up on real hardware, and how to read results within device constraints.

Venue
Procedia Computer Science (Elsevier), 2024

Highlights

  • Connects quantum annealing with ML: QBoost framing executed on D-Wave hardware with a clear experimental narrative.
  • Sits at the intersection of QC and ML—relevant as both fields converge in research and applied work.
  • Growing citation footprint, including follow-on work in Nature Portfolio journals (e.g. npj Computational Materials).
Week of electricity price data: truth (blue line) vs model prediction (orange points), Dec 24–29, 2023
Full report (PDF)
ML / Forecasting • Case Study 06

Electricity price prediction

This project forecasts electricity prices from historical series: cleaning and aligning the data, choosing and training a supervised model, then evaluating on held-out periods. The linked report documents the full pipeline—features, training setup, and metrics. The figure below is a one-week evaluation window: measured prices (“truth”) against model output on the same timeline. The model follows the strong daily cycle and tracks peaks and troughs well; predictions run slightly high at some maxima—a useful signal for calibration or ensemble refinement.

What the chart shows
Truth vs prediction over seven days (late Dec 2023): pattern match, timing, and residual bias at extremes.

Highlights

  • Side-by-side comparison of actual and forecast series on a dense time grid—easy to spot systematic error.
  • Captures recurring intraday structure rather than a single naive baseline.
  • Methodology, metrics, and discussion are in the PDF; the plot is a representative evaluation slice.
AI • Case Study 07

Agent Coach — long-memory Telegram assistant

A production Telegram bot where memory is a first class feature: goals and facts accumulate in simple files, load into every new chat, and improve after each reply. One codebase powers multiple specialist bots—same polished command surface, different expertise. Under the hood: Python with one-tap quality presets, optional live web lookup, and a backup model when traffic spikes so the chat keeps moving.

Model presets
Lean · balanced · max quality

Highlights

  • Session threads reset; long-term context does not—so coaching stays personal across weeks.
  • Shipped examples include fitness coaching and German exam prep; swap persona per deployment.

Core Architecture Philosophy

Performance-First

P99 latency is a requirement. Compute is optimized for cost and speed.

Security by Design

Security in schema, network topology, and CI/CD — not bolted on at the end.

Infinite Scalability

Stateless services and distributed state from ten users to ten million.

Ready to scale your vision?

Let's discuss production-ready architecture for your performance and security needs.

Schedule a ConsultationContact