Full-Cycle AI Technical Product Engineer. I design, evaluate, and productionize autonomous agents — from research and prototyping all the way to production and real customers. I care less about demos and more about evaluation science: the reliability work that makes autonomous systems trustworthy at scale.
Abdalla is a Full-Cycle AI Technical Product Engineer with 4+ years across software engineering, AI research, and technical education. He is currently a Senior AI Engineer at DataCamp, optimizing one of edtech's top-rated AI tutors through evaluation science and A/B testing.
As a former founding product engineer at Stakpak for a year, he operated in the heart of San Francisco's Silicon Valley ecosystem, competing head-to-head in R&D and product engineering against frontier agentic tools like Claude Code and OpenCode. He specializes in designing and optimizing agentic tools — from real-time agentic search and context engineering to coding-agent evaluation frameworks, hallucination detection, auto-prompt optimization, and intelligent retrieval systems that make AI agents more reliable and capable. Along the way he's built deep hands-on exposure to DevOps and cybersecurity — from Kubernetes and multi-platform deployment to security-hardened, self-hosted agent systems.
With 4+ years of combined experience across production AI systems, software engineering, and technical education — including a Senior Teaching Assistant role at the German University in Cairo — he has shipped and supervised work spanning ML, embedded systems, and autonomous systems. Previously the founding AI Engineer at Seqoon, he led a team shipping production RAG chatbots and a real-time Egyptian-Arabic voice assistant for real estate, and built an Islamic-books Notebook LLM (hybrid search, multi-agent summarization, podcast generation) as a freelance project.
A Mechatronics Engineering graduate (Excellent with High Honors) with an A+ thesis on Deep Reinforcement Learning for autonomous-vehicle navigation and an exchange semester at GUC Berlin. Beyond building products, he's passionate about growing AI communities — organizing hackathons, giving talks, and creating spaces where engineers learn and build together. He believes the best AI products come from engineers who care equally about the research, the craft, and the people using them.
A few things I work with often:
📍 Cairo, Egypt · remote-friendly
🎓 B.Sc. Mechatronics, GUC — 3.85/4.0, High Honors
🗣️ English · Arabic (native) · German (A2)
🎤 Lead instructor — AI Engineering Workshop Series
🪂 Off-screen: military parachuting & skydiving, squad leader
A detailed look at each role — what I built, and the impact it had.
Engineering work across retrieval, evaluation, and production agent systems.
A live search & retrieval system in Rust feeding a 24/7 autonomous DevOps coding agent — real-time web search backed by a caching layer for speed and freshness, with semantic reranking, trusted-domain filtering, and iterative LLM-based refinement to keep only the most relevant, current context in front of the agent. Benchmarked crawling & retrieval strategies (Crawl4AI, Playwright, Context7 MCP) against defined speed, freshness, recall, and precision metrics — landing at +40% recall and +55% precision with guaranteed data freshness.
A message versioning & checkpointing system for the agent that makes long-running autonomous execution reliable. I designed the checkpointing scheme over the agent's message history plus an automatic recovery-detection layer: when a violation is detected, the system rolls back automatically to the last safe checkpoint and switches the agent to recover — no human babysitting required for the agent to get itself back on track.
An auto-prompt-optimization layer built on top of the evaluation harness: it extracts the loss signal from evals, maps it back to the responsible section of the prompt, and then automatically rewrites that section in a loop — preserving current agent performance while pushing toward the global performance maximum. Inspired by DSPy and Adalflow, with textual-gradient refinement and multi-tier evaluation.
A system that analyzes production agent sessions and auto-designs reusable skills and rulebooks for the agent — codifying DevOps standard operating procedures for Kubernetes deployment, Cloudflare, Netlify, Coolify, and vLLM deployment so the agent applies proven, repeatable playbooks instead of improvising.
Automated agent-testing with full A/B cycles: dataset generation, LLM-as-judge pipelines with Langfuse observability, and sub-goal trajectory analysis over 1000+ coding sessions — surfacing failure patterns that end-to-end testing missed. Includes Stakpak's first DevOps eval dataset on the Harbor / Terminal-Bench standard.
A bilingual (AR/EN) real-estate chatbot on a full RAG pipeline (Milvus, Pinecone, Neo4j) with multi-turn LangChain/LangGraph agents; plus an Egyptian-Arabic voice AI for property sales — Twilio + LiveKit, TTS/STT, dialect fine-tuning (Whisper, XTTS, Sesame), low-latency inference via vLLM & Ollama.
The agentic AI solutions agency I founded and lead — from architecture to deployment.
Self-hosted agentic co-pilot wired into the SIEM, incident DB, and threat-intel stack; every verdict grounded in live tool calls against the client's own data. Grounded triage verdicts with evidence in under 30s (~90% faster time-to-triage) via parallel investigation — role-aware tiers (Tier 0→2, MITRE ATT&CK), strict memory isolation, redacted audit trail, human-in-the-loop.
A “NotebookLM for hadith & Islamic literature”: customizable notebook mode, full-database hybrid search, and voice-enabled reading. OCR & scraping pipelines for digitizing books; hybrid retrieval (Typesense BM25 + embeddings + Gemini long-context); a multi-agent system with agentic RAG, summarization, and podcast-script generation. Led a team of 2–3 engineers.