AI Agent Architect

Assaf Elovic Portfolio

Building Ora

Co-founder of Tavily (acquired by Nebius), now building Ora, and creator of GPT Researcher, an open-source autonomous research agent (29,000+ GitHub stars) that plans a research task, spawns sub-queries, and synthesizes cited findings; also built Tovana, a memory-management library specifically for giving agents persistent, context-aware recall across sessions.

Autonomous AgentsAgent MemoryResearch AgentsOpen Source

Location

Remote-first, in the most literal sense — no idea where.

Companies

Tavily

Skills

Autonomous AgentsAgent MemoryResearch AgentsOpen Source

Notable

  • GPT Researcher: 29,000+ GitHub stars

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A named architecture for a specific agent capability

GPT Researcher decomposes "research a topic" into sub-query generation, parallel retrieval, and cited synthesis — a specific, reusable planning pattern for open-ended information tasks, not a generic chatbot wrapper.

2

A standalone project focused only on agent memory

Building Tovana as its own library, separate from the research-agent product, treats memory as a distinct architectural problem worth solving in isolation rather than bolting ad hoc context storage onto one agent.

3

Protocol-level integration work alongside the agent itself

GPTR-MCP exposes the research agent through the Model Context Protocol, showing attention to how an agent should be composed into other systems, not just how it works standalone.

What AI Agent Architects can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • Give your agent's planning strategy a specific name and describe the steps precisely — a described pattern is more reusable than "it uses an LLM to research things."

  • If you solve agent memory or state management well, consider extracting it into its own library — it's a distinct hard problem worth solving independently of any one agent.

  • Expose your agent through a standard protocol (like MCP) so it composes into other systems, not just as a standalone product.

Projects

A research agent plus the pieces around it: an MCP server so other agents can call it, and a memory library for agents that need to remember their users. Everything here is built for agents as the customer.

Open SourceadvancedOngoing since 2023

GPT Researcher

PythonLLM providers (any)Web scrapingMCPLangGraph multi-agentFastAPI

An open-source research agent built around 'planner' and 'execution' agents. The planner writes research questions, execution agents gather and summarize sources in parallel, and a publisher assembles a cited report. A newer Deep Research mode explores subtopics in a tree, going deeper while keeping the whole topic in view.

Plan, run in parallel, then synthesize with citations is the structure of today's deep-research products, and this was an early open implementation. The community has kept it maintained since 2023.

  • Planner → parallel execution agents → publisher, with citations

  • Web and local-document research, with export to PDF and Word

  • Recursive Deep Research mode (~5 minutes per run)

ToolintermediateOpen source; split out of the main repo

GPT Researcher MCP Server

PythonFastMCPModel Context ProtocolDocker

An MCP server that gives Claude Desktop, Cursor and other MCP clients GPT Researcher's deep research, quick search and report-writing tools. The pitch: plain search tools return raw results that fill the context window, while this returns sources that are already filtered and validated.

It packages an agent as a tool for other agents, which is how agent systems are now combined, and it addresses a real cost: wasted context.

  • Deep research exposed as MCP tools for any client

  • Trades ~30s of latency for filtered, relevant sources

  • Docker-ready, with Claude Desktop config included

Open SourceintermediateOpen source; pip-installable

Tovana

PythonLLM providersLangGraphPyPI

A memory library for AI agents. Conversation events become memories, and memories become 'beliefs' that feed into the agent's reasoning and replies. You set up a MemoryManager with a provider and a business description, then pass it the ongoing user messages.

Personalization across sessions is still unsolved for most agents. The events → memories → beliefs model is a concrete design, not a wrapper around a vector store.

  • Three layers: events → memories → beliefs

  • Beliefs are injected to steer the agent's answers

  • LangGraph cookbook for a support agent with shared memory

Ready to build your portfolio?

Follow the AI Agent Architect roadmap — skills, projects, and timeline to get hired.

AI Agent Architect Roadmap