I design and ship agentic AI — tool-calling assistants, RAG pipelines with cited answers, and observable multi-agent workflows. Every project below is a live, clickable demo deployed on Vercel.
I'm Adel — an AI & Automation Engineer based in Cairo, focused on LLM systems that do more than chat: they retrieve, decide, and call tools to get work done.
I've built and deployed WhatsApp AI concierge bots, OpenAI-powered RAG pipelines, multi-agent screening systems, and n8n + GoHighLevel automation across hospitality, healthcare, e-commerce, and SaaS clients. My work pairs solid agent architecture with observability — Langfuse tracing and eval harnesses — so I can measure a pipeline and tune it: parse-failures to zero, accuracy up, cost and latency down.
I care about demos a recruiter can click. Each project ships as a real app on Vercel with a public URL.
Deployed apps with public URLs. Hover a card to scroll through the full page.
A LangGraph agent for healthcare prior-auth: retrieve → proposer → critic → judge, grounded in clinical criteria (RAG). Auto-resolves clear cases and escalates every denial and low-confidence case to a human, with a send-back self-correction loop. Every node traced in Langfuse. 55% auto-resolved at ~$0.005/decision.
Ask a data warehouse in plain English, get governed answers a CFO can trust. A governed semantic layer gives every metric one definition; metadata-RAG grounds the plan; the LLM emits a validated query spec (never raw SQL) that a compiler turns into SQL using only certified joins — structurally preventing fan-out inflation. Every answer is reconciled against an independent trusted-SQL report.
A LangGraph pipeline (router → retriever → scorer) fully instrumented with Langfuse. The case study uses traces to fix real failure modes: parse-failures 100%→0%, accuracy 83%→89%, latency −37%, cost −95%.
Source-cited Q&A over contracts, reports, and PDFs: OCR-fallback ingestion, structure-preserving chunking with page metadata, hybrid retrieval (Chroma + BM25) with reranking, page-level citations, and an anti-hallucination guard that flags low confidence instead of guessing.
Turns a messy Data Axle / ReferenceUSA export into enriched, evidence-backed leads: cleans the file, web-enriches each company (owner, email, LinkedIn, PE/franchise red flags with source URLs), classifies Pass / Needs Review / Reject with fit + confidence scores, and upserts into Airtable, deduped. Live upload → streaming enrichment.
Turns a messy trading-card scan or listing title into clean canonical card data (year, brand, set, parallel, grade) via entity resolution (blocking → scoring → threshold), then prices it from sold comps. Variant-aware: the same #280 prices as Base $41 / Silver $1,775 / Red $5,200.
An autonomous tool-calling agent that reads a guest message and takes action — answers from the knowledge base with citations, checks availability, reserves rooms, and saves contacts to a CRM, deciding which tools to call.
Upload an invoice, receipt, or ID → structured JSON with bounding boxes drawn over every detected field, confidence scores, and a baseline (OCR+regex) vs vision-language-model comparison. Works on real photographed and handwritten docs.
Custom WordPress plugins and CRM automation deployed for real hospitality & healthcare clients — lead capture, AI scoring, and conversion tools.
A custom WordPress plugin: a multi-step calculator that gives patients a realistic treatment-cost range in under 60 seconds, then captures the lead. White-labelled and deployed across multiple dental clinics (Mill Street, Tartan, Santa, MSD, City Centre).
10+ WordPress / Elementor sites for dental clinics and businesses — conversion-focused landing pages with the estimator, GoHighLevel lead forms, and booking funnels wired in. Two live examples: City Centre Dentistry and Smiles in Waterloo.
End-to-end AI concierge for a 3-property Egyptian boutique hotel collection — booking, pricing, check-in, events, and payments in Arabic & English.
Click a certificate to view it full-size.
Professional certification in building AI-powered apps with LLMs: the OpenAI API, prompt engineering, function calling, RAG, and chatbots. Credential ID AIEDA0011247893456.
Building autonomous LLM agents: tool use, agentic RAG, multi-step workflows, evaluation & observability across smolagents, LangGraph & LlamaIndex.
Chat completions, multi-turn conversations, output control (tokens/temperature), and the moderation API — hands-on in code.
Open to AI / LLM engineering roles and freelance work. The fastest way to reach me is email.
adelsherif8@gmail.com