Skip to content

About

I am a software engineer working at the intersection of AI systems and the infrastructure that makes them reliable enough to trust in production. The work I care about sits close to verification: evaluation frameworks, evidence, and the plumbing that makes both possible.

At SITA I work on the Cloud Infrastructure and Developer Productivity platform team. SITA is the technology company behind roughly 45% of global aviation operational data exchange, processing 35M+ messages daily. I contributed the Terraform modules and CI/CD pipelines for a self-hosted MCP and AI agent platform, built so agentic automation runs entirely inside enterprise cloud boundaries with no third-party data exposure. I also built retrieval-augmented generation over the Backstage service catalogue so agents can retrieve platform context on their own.

Alongside that I am co-founder and technical lead at Krynix, an independent assurance and evidence platform for AI agents. Agent decisions arrive over OpenTelemetry or a one-line HTTP contract and are held in independent, append-only, hash-chained custody, so any later edit or deletion is detectable. It sits above whatever enforcement stack you already run rather than replacing it.

My MSc in Machine Learning at University College London, finished with a Distinction. The dissertation, "Agentic Self-Correction in LLM Systems for Mathematical and Logical Reasoning", scored 87% and built evaluation frameworks measuring reasoning reliability and self-correction in production LLM systems. I completed it alongside full-time work.

A good deal of my work sits on the research side. I am interested in reinforcement learning, transformer architectures, and the cognitive-architecture questions that decide whether a language agent holds its context over a long task: how working, episodic, semantic, and procedural memory are organised, what the scaffolding around the model does, and which self-refinement loops actually improve a result rather than making it more confident. The dissertation was one attempt at that question and RefineX is its working implementation. Underneath it is ordinary deep learning practice: training, fine-tuning, and serving models in PyTorch and TensorFlow.

The through line is verification, evidence, and observability. Knowing what an agent actually did, and being able to prove it to someone who has no reason to take your word for it.

Alper Onder, software engineer, London
Agent memory layersFour memory stores read by an agent control loop: working memory, which occupies the context window, plus episodic memory of past episodes, semantic memory of facts and documents, and procedural memory of skills and tools. The lower three are retrieved into the context window rather than held in it. A self-correction path returns output to working memory.working memoryin contextepisodic memorypast episodessemantic memoryfacts, documentsprocedural memoryskills, toolscontext windowloopself-refinelower stores are retrieved into context, not held in it
Fig. 2Agent memory layers. The decomposition I keep returning to, following the cognitive-architecture framing used for language agents. Working memory occupies the context window; episodic, semantic, and procedural memory are retrieved into it. Most agent failures I have looked at are retrieval failures rather than reasoning failures.

Record

  1. Present

    Co-founder and Technical Lead, Krynix

    Independent assurance and evidence platform for AI agents.

    • Technical lead across a team of five, and the decision maker on product direction.
    • Built ingest over OpenTelemetry and a one-line HTTP contract, with append-only hash-chained custody that makes any later edit or deletion detectable.
    • Normalises agent decisions into one evidence model across frameworks, mapped toward SOC 2, the EU AI Act audit file, and FINRA retention.
    • Built a transparent MCP gateway that emits decisions with no agent-side code changes and blocks prompt-injection-driven exfiltration at the trajectory level.

  2. Aug 2025 - Present

    Associate Cloud Infrastructure Engineer, SITA

    Promoted from Graduate Software Engineer. Cloud Infrastructure and Developer Productivity, London, UK.

    • Contributed the Terraform modules and CI/CD pipelines for a self-hosted MCP and AI agent platform, now in pilot with active usage across early adopter teams.
    • Built retrieval-augmented generation over internal platform data so agents can autonomously retrieve context from the Backstage service catalogue.
    • Joined the Backstage internal developer portal as the first engineer during early-stage development, contributing to architectural decisions.
    • Helped ship a portal now surfacing 25+ cloud services, APIs, and infrastructure resources as self-service capabilities.
    • Sole owner, end to end, of an interactive engineering training platform served through Backstage TechDocs, with code labs, assessments, and progress tracking.
    • Owned the Terratest scaffolding service end to end, replacing a manual script setup with a self-service Backstage form that branches, copies test components, triggers CI, and opens a PR.

  3. Sep 2024 - Jul 2025

    Graduate Software Engineer, SITA

    Cloud Infrastructure and Developer Productivity, London, UK.

    • Led design and rollout of a department-wide testing and automation platform, adopted by 10 to 20 engineers across 5+ teams and growing toward 30+ engineers across 15+ teams.
    • Saved approximately $40k annually by removing duplicated tooling.

  4. 2024 - 2025

    MSc Machine Learning, Distinction

    University College London

    • Dissertation: "Agentic Self-Correction in LLM Systems for Mathematical and Logical Reasoning", scored 87%.
    • Built evaluation frameworks measuring reasoning reliability and self-correction in production LLM systems.
    • Completed alongside full-time work at SITA.

  5. Mar 2024 - Oct 2024

    Founder, EtzAI

    A small AI application studio delivering full-stack LLM-powered solutions. Now dormant.

  6. 2021 - 2024

    Freelance Software Engineer

    Client work via Fiverr, and full-stack systems for SIS-MAK, an industrial automation manufacturer building robotic control bands, quality-control gates, and assembly line systems for the automotive industry.

    • Real delivery under real constraints, alongside full-time study.

  7. 2019 - 2023

    BSc and MEng Computer Science with Artificial Intelligence

    University of Leeds. Four-year integrated programme.

    • Dissertation: Intelligent Robotics, First Class.