6+ years building backend that moves real money — currently shipping consumer-lending on Botim, the UAE super-app (150M+ users). Lately: GenAI / LLM apps on the side — RAG, agents, MCP.
Hover any chip for a proficiency read.
Click any card for detail.
Credit-card / BNPL hybrid built 0→live on Botim: lending engine, real-time credit-limit management, payments & reconciliation.
Owned the lending engine end-to-end — monthly statement generation, instalment recalculation, late-fee + VAT, and real-time credit-limit pushes to Botim — plus payment reconciliation and SMS/push notification hooks. Java · Spring Boot · MySQL · Apache Dubbo, across a dev → sim → UAT → prod pipeline with production hotfix support.
Salary-advance product live 8+ months on Botim: identity-verified onboarding (OTP / face), scheduled disbursement, repayment, and collections / capital-recovery.
WPS-based salary advance covering risk/eligibility checks, OTP + face-verified onboarding, scheduled auto-disbursement (with stale-application guards), repayment, and capital-recovery (collections) flows — in production for 8+ months.
Grounded, cited doc Q&A — chunking, embeddings, pgvector (HNSW) retrieval, abstention, and a faithfulness eval harness.
Self-built pipeline: chunking → embeddings → pgvector (HNSW) retrieval → grounded answers with inline citations, and abstention when the context is too weak to answer. Ships with an LLM-as-judge faithfulness eval harness. Python · FastAPI · Postgres/pgvector. Personal project on public / fictional data.
Hand-rolled tool-calling agent loop and a LangGraph multi-agent graph: intent routing → specialist nodes → tools/RAG → a KYC journey with human-in-the-loop.
A plan/act/observe loop (max-steps, error recovery, tool allow-list) plus a LangGraph graph: intent classifier → specialist nodes → tools/RAG → a KYC application journey that hands off to a human when confidence is low. Self-built, public / fictional data.
Built a Model Context Protocol server (FastMCP) exposing tools; connected and verified live in local Claude Code.
A FastMCP server exposing a small tool-set over the Model Context Protocol, connected to local Claude Code with real tool calls round-tripped and verified. Self-built.
FastAPI gateway fronting local models with PII redaction + output guardrails, intent routing (RAG vs chat), fallback, and observability/cost — the server-side control plane.
A FastAPI control plane in front of self-hosted models: PII redaction inbound and a leak-check outbound, intent routing (RAG vs chat), provider fallback, and a /stats endpoint for cost + latency observability. Self-built.
Order management, price engine, trading terminal, and market making — with real-time price-feed backends over raw sockets.
Full platform across back office, order management, a price-matching engine, the trading terminal, and market-making — with live price-feed ingestion over raw sockets and Redis-cached hot paths (~70% faster responses).
ML forecasting across financial instruments that generates actionable, data-driven trade suggestions.
ML models forecasting trend across multiple instruments and surfacing actionable trade suggestions; built the feature pipeline and the serving path.
ELT pipeline + NLP clustering of multi-source news into evolving storylines.
An ELT pipeline that ingests multi-source news and clusters it by topic (NLP) into evolving storylines you can follow over time.
Spring Boot backend for camera fleets: OTA updates, live-feed monitoring, dynamic feature deployment.
Spring Boot backend for camera fleets — device-server comms, OTA firmware/model updates, live-feed monitoring, dynamic feature deployment — plus native client libraries across Android, Raspberry Pi, and Jetson Nano.
Proposed a linear-time 2D convex-hull algorithm (O(n) vs O(n log n)) with visualizer; order-statistics set via Fenwick tree.
Proposed a linear-time O(n) construction for the 2D convex hull over integer points (vs the general O(n log n)) with a visualizer, and an order-statistics set/multiset built on a Fenwick (BIT) instead of a balanced BST. Academic work.
Schema → APIs → reconciliation → deployment → production hotfixes, across dev / sim / UAT / prod.
Daily AI-assisted development (Claude Code) integrated into a real production workflow.
Cut API latency ~70% with Redis, scaled webhook throughput ~10×, and keep money-movement flows reconciled and auditable in production.
Continuously tracking new research, tools, and performance optimizations — plus the security advisories / CVEs that matter — and folding what’s useful into how I build.
A role, a project, or a hard problem — send it over and I’ll get back to you.
[email protected]