I'm a forward deployed AI engineer. I embed with supply chain, logistics, e-commerce and operations teams to ship production AI agents, evals and automation — turning complex workflows into measurable, autonomous infrastructure.
I work the way a forward deployed engineer does: embedded with your operators and leadership to map the real workflow, then designing and shipping AI agents, LLM evals and automation that actually move P&L. My background spans computer science, business strategy, and applied research — and I've spent the last decade shipping production systems with operators who can't afford toys.
I'm founder-minded, allergic to over-engineering, and obsessed with the rough edges where models meet messy real-world ops.
End-to-end automation pipelines that wire LLMs, RPA and integrations into the systems your team already lives in.
Tool-using agents with planners, evals and guardrails — designed for narrow operational jobs, not chat demos.
Mapping, redesign and execution of cross-tool workflows — from intake to fulfillment, with audit trails operators trust.
Slotting, dispatch, ETA prediction, anomaly detection — applied AI for the people who actually move the boxes.
Internal tools, dashboards and customer surfaces built on a modern stack — Next.js, edge infra and AI-native UX.
Roadmaps, build-vs-buy, model evals, and exec readouts — translated for boards and operators, not just engineers.
Reinforcement-learning slotting engine for a 380K sqft 3PL — re-locating SKUs nightly based on velocity, affinity and pick-path cost. Cuts pick travel by 31%.
Multi-channel ads agent that pauses, scales and clones campaigns based on ROAS bands and creative fatigue.
Bid, keyword and ASIN-level optimization across SP/SB/SD with weekly playbook adjustments.
Streaming anomaly classifier across order, inventory and ops events — flagging fraud, misroutes and supplier drift.
Single pane of glass for ops leadership — agent activity, KPI swings, exception queues and decision audit, with one-click human override.
Voice + chat agent triaging service requests, dispatching field crews and predicting ETAs across a 200-vehicle fleet.
Mid-size 3PL serving DTC brands. Replaced quarterly slotting reviews with a nightly RL agent reading WMS + order velocity.
Voice + chat dispatch agent, fully integrated with field-service software. Live in 9 weeks, payback in 3 months.
Unified attribution, anomaly flags, and an agent that reallocates spend daily based on contribution margin — not ROAS theatre.
The pattern is uncannily consistent: pilots ship in two weeks, win the demo, then implode against real ops. The cause isn't the model. It's the absence of an evals-first operating model and a clear human-in-the-loop contract. Here's the playbook I run with every enterprise client.
Read essay →A short take on why operators sleep on slotting and the dollar math behind nightly autonomous re-locates.
Read essay →How to sequence platform spend so you don't pay a SaaS tax twice when the substrate shifts under you.
Read essay →If your evals don't reflect production, your agent doesn't ship. A walkthrough of the eval suite I deploy on every project.
Read essay →A forward deployed engineer is a software engineer who embeds directly with a client's team to take a real problem from discovery to production. Unlike a traditional consultant, an FDE writes and ships the code; unlike an in-house AI engineer, an FDE starts from the operation's workflow, data and constraints rather than from a product roadmap. Zach (Sadok) Abdelli works this way with operations-heavy businesses.
Zach (Sadok) Abdelli is an AI engineer and forward deployed engineer (FDE) consultant. He designs, builds and runs production AI agents, LLM evaluation suites, workflow automation and AI command centers for supply chain, logistics, e-commerce and operations teams.
Supply chain, logistics, third-party logistics (3PL), field service and e-commerce. Typical problems include warehouse slotting optimization, dispatch and ETA prediction, customer service and dispatch agents, operational anomaly detection, advertising automation across Meta, Google and Amazon, and executive AI command centers.
Most engagements run 6 to 14 weeks, embedded with the client's team. They start with an operations audit and an evaluation suite written before any agent is built, move through a shadow-mode period where the system runs alongside the existing process, and graduate authority as eval scores justify it. Every engagement ends with production code, evals and a runbook the client's operators can own.
An AI consultant typically advises on strategy and vendor choices; an AI engineer builds the system. Zach combines both: he runs roadmaps, build-vs-buy analysis and executive readouts, and he also ships the agents, integrations and evals himself.
Models from Anthropic (Claude) and OpenAI; LangGraph, the Vercel AI SDK, n8n and Make for agents and orchestration; Pinecone, Postgres, Supabase, Snowflake, dbt and Segment for data and retrieval; AWS, Cloudflare and Twilio for infrastructure; and Next.js, TypeScript and Tailwind for web apps. Model choice is treated as a vendor decision; the evaluation suite is the durable asset.
Email contact@sadokabdelli.com or use the contact form on sadokabdelli.com. He takes on a small number of engagements per quarter and responds within 24 hours.
If you run an operations-heavy business and you're tired of pilot-ware, let's talk. I take on a small number of engagements per quarter — usually 6–14 weeks, embedded.
Building in public.
Operating in private.
Most of my essays, breakdowns and ops experiments live on LinkedIn — where the operators actually read.