What I take on

I’ve built and run backend systems serving tens of millions of users — Cafebazaar’s ad platform at 51M-user scale, Divar’s Real Estate vertical from scratch — and migrated a live monolith to microservices without downtime. The work I’m best at is the kind where the headline problem looks like one thing (latency, downtime, slow team) and the actual problem is something one layer down (service boundaries, retry storms, deploy coupling).

What I look at:

• Bottlenecks under real load — DB, cache, gateway, queue
• Service boundaries — where to split, where to keep together, where the team shape is wrong for the system shape
• Failure modes — what happens when one component is slow or down
• Event-driven vs request-response trade-offs
• Observability gaps — what you can’t see is what will bite you

Typical engagement: 1 to 2 weeks of async work, 2 sync calls, written deliverables your team can act on.

I build AI features that ship — not impressive-looking demos that fall over in front of a real customer. The unsexy work is the work that matters: caching responses so cost doesn’t run away from you, falling back gracefully when the API is degraded, structuring prompts so the model can’t return something the system can’t handle, building evals so you actually know whether a change made things better.

Things I’ve shipped or am currently shipping:

• RAG over PDF / DOCX / plain text, with citation-anchored answers
• Embeddings + vector retrieval (Pinecone, self-hosted Elasticsearch)
• Multi-tenant LLM systems where tenant isolation is non-negotiable
• Agent platforms with predictability constraints — agents that you can run twice and get the same answer
• LLM integration with OpenAI, Anthropic, and open-source models

What I don’t do: train models from scratch, do data science. I take a trained model and make it a feature your customers can rely on.