Data Scientist Spring Intern— Agentic AI

Agent Design & Implementation 

• Build and iterate AI agents end-to-end — from defining goals, personas, and constraints to wiring LLM reasoning with tool execution. 

• Implement planning strategies including chain-of-thought, ReAct loops, and hierarchical task decomposition to enable agents to solve multi-step problems autonomously. 

• Design and refine system prompts, few-shot examples, and guardrails that shape agent behavior, tone, and decision-making boundaries. 

Terminal-Based AI Coding & Development 

• Work extensively inside AI-powered coding terminals (OpenCode, DeepAgent) as both a user and a builder — understanding how these tools orchestrate LLM calls, file edits, and shell commands. 

• Contribute to the development of custom coding agent workflows that automate code generation, review, refactoring, and testing tasks. 

• Evaluate and benchmark terminal agent behaviors: accuracy of code edits, hallucination rates, context utilization, and multi-file reasoning. 

Model Context Protocol (MCP) & Tool Integration 

• Build and extend MCP servers (using FastMCP and similar frameworks) that expose internal tools, databases, and APIs as structured capabilities for agents. 

• Design tool schemas, descriptions, and invocation patterns that LLMs can reliably discover and call. 

• Integrate agents with external services — REST APIs, vector stores, graph databases, and internal SDKs — through well-defined MCP interfaces. 

LLM Experimentation & Evaluation 

• Experiment with a range of open-source LLMs (Qwen, DeepSeek, finetune domain-specific models) to evaluate reasoning quality, latency, cost, and tool-use reliability. 

• Explore inference optimizations such as speculative decoding, constraint decoding, structured outputs, and router-mode orchestration. 

• Build and run evaluation pipelines to measure retrieval accuracy, tool-selection precision, hallucination rate, and end-to-end task completion. 

Memory & Retrieval Systems 

• Integrate agents with a broad ecosystem of external systems: vector stores (PgVector, Milvus), relational and graph databases, REST APIs, and internal microservices, all managed through secure, least-privilege access patterns. 

• Design and test memory architectures — short-term (conversation context), long-term (vector-stored interaction history), and episodic (task-specific recall) — to improve agent continuity and personalization.