Resume

Independent engagements building production AI for B2B SaaS clients — RAG, agent platforms, multi-tenant retrieval. The caching-and-fallbacks work that decides whether an LLM feature holds up in front of a real customer.

• Built a multi-tenant RAG system for an AI legal-tech startup — chunked and embedded PDF, DOCX, and plain-text legal documents, migrated the vector store from Pinecone to a self-hosted Elasticsearch cluster on GCP for cost and control, and shipped a chatbot that answers case questions with citations back to the source document. Also generated case timelines from the document corpus.
• Production AI / agent platforms for B2B SaaS clients (NDA). Backend, architecture, and AI work; multi-tenancy and predictability under load.
• Going deep on AI-assisted SDLC — Claude Code, agentic dev loops, evals, pressure-testing what holds up under production constraints.

Tech lead for the ad platform at Cafebazaar (Iran's largest Android store, 51M users). Owned a 10+ engineer team through a live microservices migration and the systems work that grew ad revenue 10× over two years.

• Led the ad-platform rebuild — split the legacy monolith into independently-deployable services, consolidated the service mesh so we ran fewer pieces with clearer ownership, and doubled the minimum ad price. Ad-placement and ML-model A/B tests on top of that. Net effect over two years was a 10× lift in sustainable ad revenue.
• Designed the ML fallback path. Real-time ad serving cannot wait for a slow model and cannot go down when one is degraded — built timeout-driven and degradation-driven fallback strategies that salvage revenue when ML services are slow or unavailable.
• Built the Kafka + Redis event pipeline that handles 50K events/sec in production. Idempotency keys on every event, Redis-tracked deduplication, batched DB writes — at-least-once with dedup gives us effectively exactly-once.
• Cut p99 latency on real-time ad estimation from 700ms to 25ms (same endpoint, more traffic) via multi-tier caching with TTL-based and async refresh, request deduplication, graceful fallback to heuristic estimates when the model service was degraded, and batch inference where latency allowed.
• Shipped a Video-on-Demand platform in a single quarter, coordinating across teams to get a production-ready streaming service out the door.
• Raised sprint velocity ~30% (measured against story points and OKR items delivered per quarter) by tightening Agile workflows and putting automated CI/CD checks in front of every PR.
• Mentored 7+ junior engineers; 5 promoted internally. Code-review cycle time dropped about 20% through clearer standards and better tooling.

Shipped 15+ high-reliability backend APIs at 51M-user scale, decomposed the legacy payment monolith into independently deployable services, and cut data loss on the stats aggregation service from ~15% to under 1%.

• Built and operated 15+ backend APIs sustaining 99.9% uptime across 51M users. Uptime measured per-request via Grafana dashboards over Prometheus data, 6h/1d/1w/1mo windows.
• Split the legacy payment service into 3 independently deployable microservices. Deployment-related downtime dropped about 40% once releases stopped requiring the whole monolith to redeploy.
• Reworked the stats aggregation service's recovery path. Data loss caused by infrastructure instability went from ~15% to under 1% — more than a 10× reduction, and the service could recover automatically from outages rather than needing manual intervention.
• Built an internal telemetry / distributed-tracing tool that other teams adopted for observability and debugging.

Took the Real Estate vertical at Divar (30M users) from a brand-new domain with no infrastructure to a working revenue line, then stepped into tech lead as the team grew.

• Built the Real Estate vertical from a blank slate — backend services, domain modelling, the first revenue-driving features. The team and the vertical scaled together from there.
• Stepped into Team Lead as we grew. Put Agile workflows in place and worked on the standards that let the team make decisions without me in the room.

Designed and built Divar's centralized feature-flag and A/B testing service (presented to the CTO, adopted org-wide), and shipped 50+ APIs powering the Real Estate vertical with >90% test coverage.

• Designed a company-wide feature-flag and A/B testing service. Took the proposal to the CTO; the design was approved; built and shipped the service. Enabled controlled rollouts and A/B tests across millions of users — kept as a standard part of feature delivery after I moved on.
• Shipped 50+ APIs powering the Real Estate vertical at Divar. >90% test coverage was the bar for the vertical, not just a personal one.
• Worked across product and engineering on the revenue-driving features that ran on top of those APIs.

Built fraud detection for Cafebazaar's peer-to-peer payments — rule-based scoring with prioritised review queues — that cut manual moderation workload by ~70%. Also shipped 20+ payment back-office API endpoints and a percentage-based bill-splitting system.

• Built a rule-based fraud detection system on the P2P payment platform — velocity, amount-pattern, device-fingerprint, and account-age signals → risk score → auto-decision or moderator queue. Manual review burden dropped from 100% of transactions to ~30%; moderator workload down ~70%.
• Shipped 20+ back-office REST endpoints for the payment platform.
• Built a percentage-based bill-splitting and payment flow — the kind of finance feature where the test suite earns its keep, because the edge cases (rounding, partial settlements, retries) are where money goes missing.
• Worked closely with Android and iOS teams on API contracts; built an internal moderation dashboard so the fraud team could act on the system's output.