With 10+ years of engineering experience, I take AI products from first prototype to production and paying customers, and I have been the first engineer on new products several times. At Bosch I set the architecture of a multi-agent platform and built its core agent. As Novee’s first engineer I built three LLM products, and all were sold. At Peppy I led the AI of an employee-benefits platform that was sold to enterprise clients and spun out as Euphoric. Earlier, I was the founding engineer of Supersense, an assistive app for blind users, and co-founded an ed-tech startup. Alongside engineering, I am doing a PhD on language models that learn from experience, and my open 27B document agent was a joint winner of the ICDAR 2026 DocVQA challenge.
Experience
Staff AI Engineer — Independent Contractor
Bosch Legends Lab — multi-agent platform for sourcing external experts · Jan 2026 – present
- Architected a greenfield multi-agent platform around durable workflow execution and event-driven services, so agents run for hours at scale, survive failures and resume where they stopped, a design that held up where general-purpose agent harnesses stalled on long tasks.
- Consolidated the platform’s early microservices into a monolith until service boundaries were clear, which let a small team ship its first use cases quickly, then split it into services behind a gateway.
- Built the core agent, a neuro-symbolic research agent that finds experts across the web and assesses them against a project’s criteria, in runs of hours over up to ten thousand candidates. It outperforms other deep-research tools by a large margin on our benchmarks, and recruiters prefer it to their previous tools.
- Set up the new team’s engineering culture: RFCs for design decisions, CI/CD, API and observability standards, a roadmap of small milestone releases shaped with the PM and business manager, shared coding-agent workflows, and evaluation-driven agent development.
Novee AI — LLM products for insurance underwriting · May 2025 – Dec 2025
First engineer hired; worked directly with the CTO.
- Built an LLM extraction product, delivered on a tight deadline, that won Novee a client contract: an ensemble of agents that pulls insured assets from messy PDFs, spreadsheets and email threads and reliably automates whole categories of submissions.
- Built evaluation from day one and, with the PM and labeling team, fixed unreliable labels and product assumptions before optimising, so every design choice was decided by data.
- Built a superhuman contract-comparison agent that caught material changes that underwriters missed. The labeling interface I built for its evaluation shipped to clients as the product’s interface.
- Designed self-correcting extraction agents that check their output against independent evidence in the documents and fix what fails, improving reliability substantially; other engineers reused the approach in other products.
Senior AI Software Engineer — Peppy Health
First engineer hired for AI.
- Led the AI of an employee-benefits platform, from the prototype the business team tested with clients to a product sold to enterprise clients and spun out as Euphoric.
- Shipped its benefits assistant agent, hybrid search and recommendations, choosing each algorithm by offline evaluation against expert labels.
- Led most of the platform’s backend architecture while the product was still taking shape, including a relational data model that absorbed many changes and an event-sourced Stripe checkout I built alone, safe to interrupt and resume, which other engineers extended to new product types.
- Built personalised, event-triggered email campaigns that non-technical staff run themselves, from prototype to production.
- Launched the health app’s AI chatbot globally, building its backend and AI alone from idea to production, with clinician evaluation and safety guardrails for urgent and mental-health cases.
- Mentored junior engineers through pair programming and gave the company’s first knowledge-sharing session, on information retrieval.
Senior Software Engineer — Tractable
- Led cost-aware leakage detection end to end, the capability a prospective client required for Tractable’s entry into a new European market; it turns model predictions into cost-optimal repair decisions to flag overcharging.
- Built the human-in-the-loop review platform that routes uncertain model decisions to expert reviewers.
- Brought ML models into production in a high-scale claims-processing system.
- Pushed for CI/CD, which the organisation adopted, and mentored junior engineers and researchers.
Software Development Engineer — Amazon Web Services
- Maintained and extended high-scale, low-latency pricing and offer services in AWS Billing; took part in on-call, wrote design docs and owned parts of projects.
Founding Engineer (ML and Mobile) — Mediate / Supersense
- As the first engineer at an MIT spin-off, built an on-device document-detection model end to end, from data collection and labeling to training and deployment; it guides blind users to frame documents with the camera. Also built an intent-recognition model on the camera feed and a synthetic-data pipeline for hard cases. Took the app, Supersense, from prototype to released Android and iOS apps.
Co-founder and Software Engineer — Comind
- Co-founded and built a speech-based English-learning product for children: web app, serverless backend and user testing.
R&D Engineer, Structural Dynamics — Turkish Aerospace Industries
- Built flight-test modal analysis software to validate aircraft structural models, replacing an expensive licensed tool. Wrote the department’s onboarding handbook for engineers new to aerospace.
Selected work
- ICDAR 2026 DocVQA, joint winner (8–35B category) · open 27B code agent that tells a vision model where to look · blog
- Language of Learning, first author · position paper on models that learn from experience at test time · SSRN 2026
- ArgGYM, third author · defeasible-reasoning benchmark, under review at ICLR 2027 · proposed turning the work into a benchmark, built and ran its evaluation · arXiv
- Fine-tuning vs prompting for extraction · string-match metrics and LLM-judge ratings rank the two differently · dataset · blog
- dspy-temporal, open source · durable execution for DSPy programs · GitHub
Education
PhD, Information Systems — Middle East Technical University (METU) ·
MSc, Data Informatics — METU · Thesis on knowledge-graph-augmented multi-hop question answering with LLMs.
BSc, Mechanical Engineering — METU ·
Skills
- AI
- LLM agents and multi-agent systems, RAG and hybrid retrieval, evaluation design and LLM-as-judge, document extraction, RL post-training, PyTorch, Hugging Face
- Backend
- Python, TypeScript, Java; FastAPI, workflow orchestration (Temporal), event-driven systems (Kafka, NATS), event sourcing, PostgreSQL, Redis
- Infrastructure
- AWS, Azure, GCP, Kubernetes, Docker, CI/CD, OpenTelemetry
- Languages
- Turkish (native), English (professional)