Deepseeker

💡 The Problem: When "Probably Right" Isn't Enough

Standard LLMs (like ChatGPT) are great for creative writing but dangerous for High-Stakes Research.

For financial analysts tracking real-time market events or developers debugging breaking API changes, a 99% correct answer is still a failure. When asked about specific, recent, or hyper-local topics (like "Hackathons at UIUC in Jan 2026"), standard models often:

Hallucinate plausible-sounding but fake events.
Rely on outdated training data.
Fail silently without verifying their own work.

We realized that for AI to be trusted in professional workflows, it needs to stop guessing and start verifying.

🚀 The Solution: DeepSeeker

DeepSeeker is an Agentic Research Engine that doesn't just "search"—it self-corrects.

Unlike linear RAG chains that break if one step fails, DeepSeeker uses a Recursive "Plan → Search → Critic → Retry" Loop.

Transparency by Default: Instead of a "black box" answer, we visualize the entire decision-making process so users can trust the result.
Self-Healing: If a search yields low-quality results, the Critic Agent activates, rewrites the query, and retries automatically until verified data is found.

🏗️ How we built it (Architecture)

We built a non-linear state machine using FastAPI (Python) for the brain and Lovable (React) for the frontend.

🧠 The Core Loop

Planner Agent: Breaks the user query into research steps.
Worker Agent: Executes searches using Tavily API.
Critic Agent (The Novelty): Evaluates the search results. If the data is poor, it triggers a Retry Loop with a refined prompt.
Reporter Agent: Synthesizes the final answer with inline citations pointing to real URLs.

🛠️ Sponsor Integrations (Deep Dive)

1. Keywords AI (Core Infrastructure)

This project would not be possible without Keywords AI. We used it as our Full-Stack Observability & Control Plane:

The Gateway: We routed all LLM requests through the Keywords AI Gateway (api.keywordsai.co).
Traces: This was critical. We used Traces to debug our recursive loop. We could visually see the exact moment the Critic Agent rejected a result and triggered a retry, allowing us to optimize latency.
Prompt Management: We fully decoupled our logic. The Planner, Critic, and Reporter prompts are all managed via the Keywords AI Dashboard, allowing us to iterate on the "Agent Persona" without redeploying code.

2. Lovable (UI/UX)

We used Lovable to rapidly build a "DeepSeek R1-style" interface. It features a complex "Live Agent State" timeline that visualizes the backend WebSocket states (Planning, Searching, Retrying) in real-time, turning the wait time into an engaging experience.

3. Supabase (Database)

We used Supabase to archive research sessions. We store the full JSON logs of the agent's thought process, allowing users to review past research.

4. Trae (AI IDE)

We utilized Trae as our intelligent development environment. Its AI coding assistance accelerated the build of our FastAPI backend and complex agentic logic, letting us focus on the recursive architecture instead of boilerplate code.

⚠️ Challenges we ran into

The hardest part was tuning the Critic Agent. Initially, it was too lenient (accepting bad data) or too strict (looping forever). We used Keywords AI Traces to analyze the Critic's decision-making process and tweaked the system prompt in the dashboard until we achieved a stable "Self-Healing" balance.

🏅 Accomplishments that we're proud of

Building a Self-Healing architecture that actually works. Seeing the UI turn from "Warning" (Red) to "Success" (Green) without human intervention is magical.
Achieving full observability via Keywords AI.
Creating a polished, production-grade UI in under 48 hours.