Aletheia

Aletheia, a proactive AI research agent, monitors topics 24/7, filters noise, and turns complex papers into instant podcasts, videos, and briefings—so you learn faster and never miss a breakthrough.

Comment

Inspiration

In the fast-paced world of AI and technology, new research papers and breakthroughs are published every hour. Researchers, students, and professionals often feel overwhelmed by the sheer volume of information ("FOMO"). We wanted to build a tool that doesn't just wait for you to search, but proactively works for you—filtering the noise and transforming dense academic text into engaging formats like podcasts and video lectures that fit into a busy lifestyle.

What it does

Aletheia is an autonomous research agent platform.

  1. Research Radars: Users define topics (e.g., "LLM Agents," "Quantum Computing"). Aletheia monitors sources like Arxiv and Twitter/X 24/7, intelligently ranks findings based on semantic relevance and your historical feedback.
  2. Proactive Multimodal Briefings: Periodically, Aletheia Radar delivers a curated briefing of each top-ranked paper in user's preferred format - audio podcasts, video lectures, or text digests.
  3. Exploration: Don't just read—chat with your research. Dive deeper into any paper from your feed, ask complex questions, and save multimedia notes as you learn.
  4. Collaborative Projects: Turn insights into output. Partner with Aletheia on long-term projects to co-author papers, generate slide decks, or produce educational content tailored to your needs.

How we built it

  • Backend: Python with FastAPI and Google Agent Development Kit (ADK). We used the ADK to orchestrate a multi-agent system (Root Agent delegating to Specialist Agents.
  • AI Models: gemini-3.0-pro handles the heavy lifting—summarization, semantic ranking, and even generating the scripts for podcasts and slides.
  • Frontend: A modern React application (Vite + Tailwind CSS) providing a dashboard for managing Radars and viewing multimedia content.
  • Database: Firestore for storing user preferences, radar configurations, and research history.
  • Scheduling: APScheduler manages the background tasks that run the "Radars" hourly, daily, or weekly.

Challenges we ran into

  • Complex Agent Orchestration: Coordinating multiple specialized agents (e.g., separate agents for search, synthesis, and media generation) was non-trivial. We used the ADK pattern to to strictly define agent scopes and manage handoffs, preventing loops and confusion
  • Hallucination in Citations: Early versions would sometimes invent papers. We solved this by grounding agents with restrictive tools (like search_arxiv) and enforcing a verification step where the LLM must validate the source URL against the retrieved metadata.
  • Handling Large Contexts: Enabling Q&A across multiple full-text PDFs required careful token management and context window optimization to maintain performance without losing details.
  • Real-Time Feed Ranking: Hot topics can generate hundreds of preprints and tweets daily. To surface the most relevant content, we implemented semantic ranking combined with collaborative filtering. As our user data grows, we plan to refine this algorithm further.
  • Session & State Management: Managing long-running research workflows across different stages (Radar, Exploration, Projects) introduced complexity in maintaining chat context and database consistency. We unified the session handling to ensure documents and context flow seamlessly between stages.
  • Intuitive UX for Complex Workflows: Designing a UI that balances power with simplicity was challenging. We conducted intensive user interviews to refine the interface, ensuring it remains accessible without sacrificing the depth needed for serious research.
  • Multimodal Synthesis: Generating high-fidelity audio podcasts and video lectures from technical papers was extremely challenging. We had to carefully prompt the models to adopt a natural, conversational tone for the audio while simultaneously ensuring the visual content in slides accurately represented the complex data, all synchronized in real-time.

Accomplishments that we're proud of

  • True Multimodality: We successfully built a pipeline that turns dry academic PDFs into engaging audio podcasts and video lectures. Hearing a research paper summarized as a lively conversation is a game-changer for accessibility.
  • Semantic Ranking at Scale: Implementing a custom relevance engine that combines LLM reasoning with collaborative filtering signals, ensuring users see the most impactful papers first, not just the newest.
  • Clean, Scalable Architecture: Refactoring the complex multi-agent system into modular services (Radar, Exploration, Projects) using the ADK pattern, making it robust enough for real-world usage.
  • User-Centric Design: We iterated through multiple designs to create an interface that simplifies complex research workflows without dumbing them down.

What we learned

  • The Power of Verification: Hallucinations can be drastically reduced by forcing agents to strictly validate URLs and metadata before presenting them to the user.
  • Workflow Transformation Depends on UX: Introducing an AI-driven, proactive research workflow is a significant shift from traditional manual searching. User adoption hinges entirely on how intuitive and trusted the interface feels. If the AI acts as a "black box," users revert to old habits. Transparency and clear feedback loops were essential.
  • Context Management is Critical: Simply feeding all histories, sometimes including multiple PDFs, into an LLM context window is expensive and slow. We learned to optimize by extracting only relevant sections and using efficient summarization chains.

What's next for Aletheia

  • Mobile App: Enable push notifications for radar findings so users can listen to their daily briefing on the go.
  • Expanded Data Sources: Integrate Twitter/X, GitHub trending repositories and customized tech blogs into the Radar system for a holistic view of the tech landscape.
  • Personal Knowledge Base: Develop a comprehensive, interconnected repository where every paper read, every note taken, and every insight generated is indexed and easily retrieval, acting as a second brain for the researcher.
  • Collaborative Research: Allow teams to share Radars, co-annotate papers, multi-media outputs, and build shared knowledge bases.
  • Beta Testing & Commercialization: We plan to onboard a pilot group of individual researchers to iterate on existing features based on deep feedback. Parallel to this, we will develop a Freemium B2C model and explore B2B enterprise features for R&D teams.

Built With

Share this project:

Updates