Kagan Demirhindi

More than 14,000 companies use Pendo to create better software experiences, including Salesforce, Cisco, Verizon, HubSpot and Zendesk.

• Took Pendo's AI agent, the AI layer of a platform 14,000+ companies use, from pre-launch to production, building multi-agent orchestration, skills, prompt and context engineering, tool calling, MCP tooling, knowledge retrieval and semantic / vector search. Public adopters include Datasite and Meevo.

• Proposed and built Pendo’s AI evaluation tool (Python, React, PostgreSQL): LLM-as-a-Judge verdicts, repeatable test suites, repeated runs with pass thresholds for non-determinism, and comparison across agent harnesses including MCP and LangGraph setups. Engineers, PMs and managers now use it as an evaluation gate for model, MCP and launch decisions.

• Expanded the agent into Japan, from raising the opportunity to owning the regional deployment architecture and delivery required under data residency rules, run end to end with Legal and Platform Ops.

• Cut the agent’s error and issue rates in three months by debugging production traces and user feedback, and adding guardrails.

• Reduced unsupported requests by shipping a knowledge-base retrieval (RAG) tool, after production traces showed that customers’ how-to questions were going unanswered.

• Worked with four teams for weeks or months at a time, showing on their own real tasks how going AI-first makes them both faster and more reliable, and raised the AI adoption their managers track. Contributed to three of Pendo’s company-wide AI initiatives: its AI agent, its MCP server and the AI-native rebuild of Pendo Feedback. Named Pendo's AI MVP for Q2 2026.

The only project management platform built specifically for client work.

• Designed and built an AI onboarding flow that generated customised project plans, prioritised tasks and assigned work from user input, doubling onboarding completion rates.

• Built Teamwork’s AI Profitability Forecaster, turning logged time, cost and billable rates into one-click revenue, cost and profit projections with confidence ranges. Self-hosted the models on AWS and selected between TinyTimeMixer and Prophet based on how much history each site had.

• Built product end to end in Vue.js, Go and PostgreSQL, using LaunchDarkly to test changes and Pendo and HubSpot to measure results before rollout.

London-based fintech startup. ePOS app that turns tablets and smartphones into powerful cash registers.

• Built Go and gRPC services on Google Kubernetes Engine and the React and TypeScript POS app they supported, including native plugins that connected phones to the bank’s payment terminals.

• Joined at the zero-to-one stage and built the company’s first merchant onboarding integration with its partner bank.

Empowering Futures with AI Innovation.

• Built backend microservices and REST APIs in Java with PostgreSQL and GraphQL, containerised with Docker, and automated build and deployment pipelines.

• Built large-scale web-scraping pipelines that collected gigabytes of data daily, then cleaned and prepared the data in Python and Pandas to train machine-learning models.

• Performed exploratory data analysis on large industrial datasets using Python, Elasticsearch and Kibana to support researchers.

Open-source skills for coding agents (Claude Code, Codex and others), each shipped with its own eval suite covering behavioural patterns, safety boundaries and baseline comparison.

senior-review settles architecture before any line-level review begins, cutting the nitpicks agents usually produce. agent-fix-loop uses repeated, trace-verified tests to improve agent reliability. chronos, pushback, shipit, socratic and timescale cover wall-clock time awareness, productive disagreement, git/PR workflows, structured thinking and AI-native delivery estimates.

Author of “Machine Learning and Its Use in Health Research”, a chapter covering machine learning, generative AI and Python applications in health research.

“How to Train Your Coding Agent”: Explored the current state of LLM capabilities along with practical patterns for managing context windows, subagents, skills and agentic workflows (bash loops) for AI-first development.

“Fewer Juniors Today = Fewer Seniors Tomorrow”: A post on the long-term talent-pipeline risks of AI hype in software engineering, viewed 2M+ times across LinkedIn and Reddit.

Graduated with a 3.11/4.0 GPA while working as a software engineer from my second year alongside coursework. Through the Ege Entrepreneurship Society, organised company visits and campus events that connected students with industry.