Andranik Khachatryan

I built an AI agent that reduces marketing performance investigation time from hours to minutes by automating root-cause analysis. When analysts notice unexpected drops or spikes in metrics, the agent investigates by forming hypotheses, executing data warehouse queries, and iteratively narrowing down to the root cause.

The system uses Google ADK with Gemini, backed by a knowledge base of 1,600+ metrics and investigation strategies. I implemented streaming responses via FastAPI, an async job system with PostgreSQL for long-running investigations, and parallel query execution for testing multiple hypotheses concurrently. Prompt content and sample investigations are version-controlled, with documented ground-truth cases for concurrently testing regression hypotheses. This reduces manual investigation time from hours to minutes.

I built a natural-language-to-query translation system that converts English business questions into structured analytical queries. Used by business users for conversational data exploration and by other AI agents for query generation.

I designed a 6-step LLM pipeline with validation at each stage: intent parsing, column discovery from curated and full catalogs, candidate reduction, query generation, and result analysis. I also implemented ground-truth verification that compares generated queries against known-good answers, tracking per-step success rates and execution timing. Features an extensible LLM provider abstraction with per-step model selection, conditional pipeline execution that skips unnecessary steps, and Redis pub/sub for real-time progress tracking.

I built an end-to-end machine learning platform that orchestrates model training, deployment, and monitoring on Amazon SageMaker. The system provides a RESTful API abstracting SageMaker's complexity into clean workflows: model registration with semantic versioning, dataset creation from data warehouse queries, two-stage async training pipelines (preprocessing + XGBoost), and two-model endpoint deployment.

I implemented MLflow integration for experiment tracking with per-iteration metrics logged from inside SageMaker containers. Features non-blocking job orchestration with client-driven status polling, temporal train/test splitting to prevent data leakage, soft-delete patterns for audit trails, and production version safety constraints. The solution is built using FastAPI, PostgreSQL with asynchronous SQLAlchemy, and Amazon SageMaker.

PyTorch implementation of paper "HiDDeN: Hiding Data With Deep Networks" by Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei: https://arxiv.org/abs/1807.09937

Image search database which indexes tens of millions of images and is able to find the original photo, even if query image is heavily cropped, resized, down-sampled, collage version of the original. Horizontally scalable, cloud-based.