Résumé

Experience

1. Graduate Engineer Trainee

   Jun 2026 – Present

   Multi Commodity Exchange of India (MCX) · Mumbai, India

2. Backend Developer Intern

   Dec 2024 – Mar 2025

   FyreGig · Mumbai, India (Remote)

   • Designed a Kafka-based real-time broadcast system supporting 10M+ concurrent users through dynamic topic filtering, cutting database write load by 70%.
   • Built a change-data-capture ETL pipeline with Debezium and Apache Kafka, replicating 1M+ records from MySQL to ClickHouse in under a second at 99.9% consistency.
   • Cut organisation member tree-view API latency by 85% (10s to 1.5s) by replacing recursive queries with breadth-first traversal, batch processing and concurrent retrieval.

Research

• Interpretable Instance Segmentation for Camouflaged Wildlife Detection

  International Conference on Advances and Developments in Computational Intelligence Systems (ADCIS 2025) · Best Presentation

• AI-Based Synchronized Video Dubbing

  International Conference on ICT for Sustainable Development (ICT4SD 2025) · Springer

Selected projects

• Sentinel: an autonomous multi-agent AI system

  Owais Barkati built Sentinel, a multi-agent orchestration framework on LangGraph and FastAPI that classifies a natural-language query's intent and routes it to a specialist agent — SQL, forecasting, retrieval-augmented generation or web search. It includes a privacy-preserving federated learning pipeline using Flower and QLoRA that fine-tunes Llama 3 and Qwen across distributed client nodes without centralising sensitive data.

• Broadcasting to 10 million concurrent users

  Owais Barkati designed and implemented a high-throughput real-time broadcast and notification system on Apache Kafka at FyreGig, supporting more than 10 million concurrent users through dynamic topic filtering. Moving fan-out from the database onto the event stream cut database write load by 70%, while TTL policies and log compaction kept message durability intact.

• Sub-second MySQL to ClickHouse replication

  Owais Barkati built a change-data-capture ETL pipeline using Debezium and Apache Kafka that replicated more than 1 million records from MySQL to ClickHouse with sub-second latency and 99.9% data consistency. The pipeline powers real-time business intelligence dashboards used for loan risk analysis, replacing batch extracts that could not keep up.

• Cutting an API from 10 seconds to 1.5

  Owais Barkati refactored an organisation member tree-view REST API at FyreGig, reducing response latency by 85% — from 10 seconds to 1.5. The fix replaced naive recursive per-node queries with a breadth-first traversal that batches each level into a single query and retrieves data concurrently, turning a request whose cost scaled with the number of employees into one that scales with the depth of the hierarchy.

• The assistant on this site has no vector database

  Owais Barkati built the AI assistant embedded in this portfolio. It answers only from the site's own case studies, publications and posts, and cites the page each claim came from. It has no vector database and performs no retrieval: the entire corpus is roughly 8,800 tokens, so it is sent in full behind an Anthropic prompt-cache breakpoint. A cached turn costs about $0.0012.

Education

Technical skills

Languages

Python, C++, C, Java, JavaScript, TypeScript, SQL

AI & Machine Learning

PyTorch, Hugging Face, LangChain, LangGraph, Retrieval-Augmented Generation (RAG), Federated Learning, QLoRA, FAISS, YOLOv8, Instance Segmentation, Explainable AI (XAI)

Backend & Frameworks

FastAPI, Node.js, Spring Boot, Flask, REST APIs, Microservices

Data & Streaming

Apache Kafka, Change Data Capture (CDC), Debezium, ClickHouse, Apache Spark, Hadoop

Databases

MySQL, PostgreSQL, MongoDB

Cloud & DevOps

Docker, Kubernetes, AWS, CI/CD, Power BI