Mark – Python, SQL, AWS

Developed and deployed a fine-tuned automated content moderation system designed to make political speech moderation more consistent, transparent, and less partisan. Combines LLM-based text analysis, fine-tuned local models, human-in-the-loop labeling, and supervised classifiers trained on stance-invariant moderation decisions to predict whether posts should be kept, removed, or down-ranked. Deployed on Modal via vLLM.

• Was a tech lead on the project;
• Fine-tuned Qwen models for custom reward modeling;
• Fine-tuned and deployed multiple supervised classifiers for performing textual analysis;
• Designed as batch content moderation system, deployed on AWS using a microservice architecture, with the fine-tuned models deployed using vLLM in Modal.

An AI agent social simulation engine for modeling how communities behave on social platforms under different algorithms and interaction rules. The system ingests social-network datasets, initializes simulated users as agents, and runs them through discrete time steps where they receive feeds, take actions, and change the platform state. It is designed around configurable runs, task-based agent execution, persistent simulation state, and metrics pipelines for comparing outcomes across experiments.

• Served as tech lead;
• Implemented state execution engine, combining deterministic rules with LLM-based intelligence for resolving agent actions;
• Added telemetry using OpenTelemetry + Opik. Added tracing and included evaluation workflows and golden datasets;
• Created scalable plugin architecture for allowing dynamic social network evolution. Designed Postgres table schemas and Neo4j graph database storage for social network analysis.

A scalable system was developed to enable natural language querying over large-scale tweet datasets (~2TB) stored in Parquet format. The solution empowered non-technical users, such as researchers and analysts, to run flexible, human-language queries like “What are the top hashtags about inflation last month?” and receive accurate, explainable results. Leveraging LLMs to translate natural language into SQL, the system executed these queries over the database and returned well-structured responses, streamlining data access and analysis.

• Designed and implemented a stateless query pipeline over partitioned tweet data using DuckDB and GPT-based SQL generation;
• Led architectural decisions for scaling from 12GB to over 1TB of data with under-1-minute latency constraints;
• Developed per-day sharded aggregation pattern and materialized view strategy to support real-time analysis;
• Built prompt-engineered GPT workflows for robust NL-to-SQL translation, with query fallback and validation;
• Integrated caching, semantic filtering, and keyword indexing to reduce scan costs and improve UX;
• Enabled cross-functional stakeholders to query and explore social media data without writing code.

A custom AI solution was developed to analyze sentiment across large-scale social media datasets, such as Twitter and Reddit. The project involved training advanced sentiment models tailored to domain-specific corpora, enabling the extraction of nuanced signals and attributes from online conversations. These models were deployed across tens of millions of posts to analyze sentiment trends over time. A novel data augmentation method was implemented by combining gold-standard labels with large language models (LLMs) to improve model performance and generalization. This system provided valuable insights into evolving public sentiment and discourse.

• Led model architecture development and model training;
• Created a custom data augmentation method using a combination of gold labels and LLMs;
• Made a custom retraining scheme and schedule to prevent model drift;
• Added proper MLOps monitoring and telemetry;
• Deployed model to analyze sentiment across tens of millions of social media conversations across various platforms.

A RAG-based application was developed to identify and surface political bias in both published news articles and social media content. The system embedded and classified tens of thousands of articles from a wide range of U.S. news outlets, analyzing political bias across multiple dimensions validated by academic research. It enabled the tracking of bias and sentiment trends over time. The solution was powered by agentic infrastructure that autonomously analyzed, aggregated, and synthesized relevant data sources to generate comprehensive and explainable insights.

• Created end-to-end RAG application;
• Designed an AWS-powered data ingestion and processing pipeline, responsible for sourcing information from dozens of news sources;
• Created agentic infrastructure to analyze, synthesize, and aggregate various data sources over time.

A modular, production-grade ML pipeline designed to process multimodal communication data such as Zoom transcripts and emails. The system performs low-latency inference using machine learning models to generate actionable follow-up suggestions and sentiment insights for end users. It incorporates AI-powered agentic components across various communication platforms to provide a unified and intelligent view of user interactions.

• Designed and deployed a scalable ML pipeline integrating Zoom and email data into contextual AI suggestions;
• Developed models for sentiment classification and follow-up generation
• Built and managed a model serving infrastructure;
• Implemented Redis and S3-based hybrid feature store with online/offline sync, using Feast for feature management;
• Tracked and versioned models and metrics with MLflow, enabling reproducible experiments and staged deployment;
• Orchestrated ingestion, preprocessing, and inference via AWS Step Functions and Prefect;
• Applied Terraform for infrastructure-as-code and modular deployment of all components.