Senior Software Engineer · 10+ Years · Istanbul, Turkey
Applying for: Systems Engineer, Workflows · ETI Team, Lisbon
About Me
Senior Data Science Engineering at Countly (1B+ daily data points). Building AI Agent CEE, RAG pipelines on Cloudflare Workers, and AI-powered analytics features.
claude-context-mode
Context-saving MCP server for Claude Code
Inspired by Cloudflare. Cloudflare's Code Mode compresses MCP tool definitions from millions of tokens to ~1K. I applied the same principle to the other direction — tool outputs. 315 KB raw output becomes 5.4 KB (98% reduction), extending session time from ~30 min to ~3 hours.
AI + Workflows
At Countly, I built a Gemini RAG pipeline on Cloudflare Workers with ~0ms cold starts. These multi-step AI pipelines are the patterns Workflows orchestrates well.
MCP Directory (CF Pages + Workers + Durable Objects) and seclawai both use multi-step tool chains with retries and error recovery — the same patterns Workflows already supports.
claude-context-mode (400+ stars) was directly inspired by Cloudflare's Code Mode. I think about developer experience as a product — and I'd enjoy contributing to Workflows' DX story around types, debugging, and testing.
Workers, Durable Objects, D1, R2, Workers AI — these aren't resume keywords, they're my daily tools. I know where the rough edges are because I ship production systems on them.
Experience
2025 — Now
Remote, Istanbul
Senior Data Science Engineering · Product development at a privacy-focused analytics platform processing 1B+ daily data points for 1000+ enterprise customers. Built Gemini RAG on Cloudflare Workers with ~0ms cold starts. Currently developing AI Agent CEE.
2024 — 2025
Remote, Istanbul
Senior Consultant · Led an 8-person engineering team across 3 products. Owned architecture decisions for EV Clearing House (OCPI 2.2.1, 50+ operators) and D-Charge (1000+ stations). Designed reusable boilerplates adopted by 5+ teams across the organization.
Previous
Remote, Istanbul
Planhat · Stockholm
Senior SWE · B2B customer success platform. Feature development and frontend architecture.
Qooper · Chicago
Tech Lead · Ruby-to-Node migration (50% faster API), unified mobile to React Native.
TDSmaker · Istanbul
SWE · 90%+ test coverage via TDD, 50% faster deployments, platform for 500+ users.
As a User & Builder
Public bucket emulation doesn't work locally after Wrangler 3. Community's most-requested fix.
I use R2 in production for my RAG system. Local dev parity is essential — deploying just to test file serving breaks the fast iteration loop.
vitest-pool-workers needs Vitest v4 support. The community is eager for a modern testing story.
I follow this closely in the community. A solid testing story is key for Workflows adoption — orchestrations need reliable test infrastructure.
ctx.id.name has a gap between types and runtime. 61 developers are asking for a seamless fix.
I use Durable Objects for real-time state in production. Good to see the fix merged — a clean identity API removes unnecessary workarounds.
Same-name columns in JOINs need disambiguation. Predictable query results build developer trust.
I use D1 for persistent state. Silent data issues in JOINs are the hardest to catch — glad to see partial fixes landing.
Bundle size affects TTFB even for unused code paths. Opportunity for smarter tree-shaking at the isolate level.
I target sub-100ms responses on Workers. Bundle size affecting TTFB even for unused paths is something I actively work around.
Official types in @cloudflare/workers-types would complete the DX circle for a TypeScript-first product.
As a TypeScript-first developer, official types are the API contract. I'd love to help ship this — it's foundational for Workflows DX.
Open Source
Effect-style composable workflow primitives for Cloudflare Workflows. Declarative DSL that compiles to CF Workflows API calls.
step().retry().timeout().compensate()
— chainable DSL
parallel()
— child DO isolation with self-spawn pattern
compile() → execute()
— validation + CF API translation
# deployed & tested on CF Workers
$ curl -X POST .../notify -d '{"userId":"u1"}'
✓ status: "complete"
✓ 3 channels notified (email, sms, crm)
✓ parent cpuTime: 0 (hibernated at $0)
✓ each child ran in separate DO
Promise.all — single DO
128 MB shared memory, 5 min shared CPU. Two 80 MB tasks = OOM crash. One failure kills all branches.
parallel() — child DOs
128 MB each, 5 min CPU each, per-branch retry, isolated failure. Parent hibernates at $0.
Why this matters: I didn't just read the Workflows docs — I built a production library on top of it. Understanding a system deeply enough to abstract it into a clean DSL requires knowing its internals: Durable Object lifecycle, step caching semantics, and the self-spawn pattern for parallel isolation.
Looking forward to discussing Workflows and the ETI team's vision.
Mert Koseoğlu · bm.ksglu@gmail.com · mksg.lu