ResearchPilot

An autonomous multi-agent deep research engine that plans, searches, fact-checks claims, and synthesizes publication-grade scientific dossiers with live DOI audit trails.

Comment

💡 Inspiration

Conducting exhaustive, publication-grade scientific and technological research currently takes domain experts dozens of hours of manual literature review across fragmented academic journals, industry reports, and preprint servers. Standard LLMs often suffer from three critical flaws:

  1. Hallucinations & Stale Citations: Generating fabricated links or unverified facts.
  2. Lack of Provenance: Inability to audit the exact sentence and DOI publication where a claim originated.
  3. Absence of Adversarial Verification: Blindly believing marketing claims rather than cross-examining competing benchmarks.

We built ResearchPilot to solve this—an enterprise-grade autonomous deep research system powered by an 8-agent state machine that independently plans, retrieves, extracts atomic claims, performs adversarial cross-examination, and synthesizes publication-ready intelligence dossiers.

🚀 What It Does

  • Autonomous Goal Decomposition: Decomposes complex scientific inquiries into multidimensional research hypotheses.
  • Academic & Web Retrieval: Executes recursive queries across Nature, arXiv, ScienceDirect, PubMed, and BloombergNEF.
  • Atomic Claim Extraction: Extracts specific empirical metrics (e.g., energy density $\text{Wh/kg}$, plasma gain $Q_{\text{plasma}} \ge 1.0$, levelized cost $\$ /\text{tCO}_2$).
  • Adversarial Verification Gate: Flags every claim as SUPPORTED, CONTRADICTED, or UNVERIFIED with confidence scores ($0.0 - 1.0$).
  • Clickable Provenance Inspector: In the final report, clicking any ✨ [Claim claim_01] tag opens an interactive side-panel displaying the exact extracted snippet, verifier reasoning, and direct external DOI citation.
  • Zero-Trust Security Sandbox: Built-in SSRF DNS firewall (blocking private IP and cloud metadata access), prompt injection quarantine blocks (<untrusted_evidence>), and runtime bound constraints.

🛠️ How We Built It

  • Multi-Agent Orchestrator: Built with Node.js 20 and Google ADK (Agent Development Kit), orchestrating 8 specialized agents (Manager, Planner, Researcher, Academic Retriever, Claim Extractor, Adversarial Verifier, Synthesizer, Critic).
  • Deterministic State Machine: A 10-stage monotonic execution pipeline (QUEUED $\rightarrow$ PLANNING $\rightarrow$ RESEARCHING $\rightarrow$ EXTRACTING $\rightarrow$ VERIFYING $\rightarrow$ SYNTHESIZING $\rightarrow$ CRITIC_REVIEW $\rightarrow$ COMPLETED).
  • Frontend Command Center: Built on Next.js 14 (App Router), Tailwind CSS, and Framer Motion, featuring real-time Server-Sent Events (SSE) streaming live telemetry and decisions.
  • Vector Embeddings & Storage: PostgreSQL with pgvector and Prisma ORM for tenant-isolated semantic knowledge graphs.

🧗 Challenges We Ran Into

  1. Preventing LLM Confirmation Bias: Initial single-agent loops frequently accepted biased marketing claims. We resolved this by creating an Adversarial Verifier & Critic Gate that forces bidirectional cross-source disputation.
  2. Live Telemetry & Monotonic State Sync: Ensuring that multi-step agent thought streams smoothly stream to client browsers without race conditions or memory leaks via persistent SSE channels.
  3. Security Against Indirect Prompt Injection: Neutralizing malicious prompt injections hiding inside scraped web text by wrapping all untrusted inputs in strict <untrusted_evidence> quarantine delimiters.

🏆 Accomplishments That We're Proud Of

  • Achieved 100% verifiable citation audit trails where every single synthesized paragraph maps back to authentic DOI literature.
  • Engineered a sleek, glassmorphic dark-mode interface with zero latency and interactive claim inspection.
  • Zero-dependency deterministic fallback mode allowing safe, instant demonstration without cloud friction.

📚 What We Learned

  • Multi-agent architectures yield significantly higher factual accuracy when agents operate with distinct adversarial incentives (e.g., Researcher vs. Critic).
  • Strict schema validation and atomic fact chunking are essential for preventing context window degradation during deep research.

🔮 What's Next for ResearchPilot

  • Integrating multi-modal PDF chart and chemical structure extraction.
  • Direct export of synthesized intelligence dossiers into LaTeX and IEEE conference paper templates.
  • Autonomous agentic webhook alerts for newly published preprints on user-tracked research topics.

Built With

Share this project:

Updates