NEXUS — Autonomous Workflow Intelligence

An AI workflow engine that executes tasks, verifies results, and recovers from failures.

Comment

Inspiration

Today’s autonomous AI research agents and search assistants suffer from two fatal flaws: hallucinated verification and silent failure.

When tasked with technical market intelligence, existing tools routinely scrape an off-topic academic preprint or blog, extract disconnected metrics, paste them into a summary, and confidently declare:

VERIFIED (All Claims Corroborated by Primary Sources)

In reality, neither the sources nor the numbers support the claims. Furthermore, when search or extraction tasks fail, agents either crash or silently hallucinate plausible answers. We built NEXUS (Autonomous Workflow Intelligence) to solve this: a closed-loop, fail-closed autonomous research engine where no claim is certified without traceable, primary-source empirical evidence, and where autonomous agents self-heal from failures without drifting off-topic.

What it does

NEXUS turns high-level research goals into verified, publication-grade intelligence briefings through an autonomous Plan ➔ Execute ➔ Observe ➔ Verify ➔ Recover ➔ Vault loop:

Autonomous DAG Planning: Decomposes user goals into parallel dependency graphs across local, API, and web tools. Fail-Closed Evidence Gate: Rejects irrelevant papers, tools, and superficial keyword matches. Separates Primary sources (official specs, original IDC/Canalys data) from Secondary reporting (blogs, reviews). Returns VERIFIED only if 100% of claims have primary corroboration; otherwise fails safely to NEEDS_REVIEW. Temporal Auditing: Distinguishes historical baselines (e.g., 2024 shipments) from current 2025–2026 developments without confusing time periods. Self-Healing Recovery: Employs Gemini Flash to diagnose failed tasks and retry autonomously without drifting off-topic. Obsidian Brain & DOCX Reports: Generates formatted Word briefings with comparison tables and syncs structured Markdown notes into a persistent Obsidian vault.

How we built it

Backend: FastAPI & Python 3.14 with SQLite state persistence. AI Models: Groq & OpenRouter: Instant task decomposition and initial DAG planning. Gemini Flash Pool (gemini-3.8-flash & gemini-3.6-flash): Semantic triage, autonomous recovery, and claim verification. Web & Search: Tavily Search with canonical URL deduplication and tracking scrubbers. Frontend: Lightweight React 19 + Vite dashboard with real-time SSE execution streaming and an interactive Obsidian Brain drawer. Document Generator: Custom python-docx engine generating styled technical briefings with quarantined evidence sections.

Challenges we ran into

False Verification: Early versions marked reports VERIFIED despite pulling irrelevant microcontroller papers. We built deterministic domain filters and a persistent rejection store. Topic Drift During Recovery: Agents fixing failed extraction tasks often drifted to broader topics. We enforced strict anchor boundaries. Temporal Conflation: Preventing LLMs from presenting 2024 historical statistics as 2026 market leaders. Bundle Bloat: Replaced a 500+ MB Next.js setup with Vite + React 19 to shrink the entire repo to 97 MB.

Accomplishments that we're proud of

10/10 Regression Test Suite: 100% passing automated tests verifying evidence integrity, deduplication, and fail-closed gates. True Fail-Closed Behavior: Transparently labels incomplete reports NEEDS_REVIEW instead of hallucinating certification. Autonomous Self-Healing: Recovers from scraper blocks and failed fetches without human intervention. Sub-100 MB Full-Stack Architecture: Complete agent, doc engine, vault, and UI in a lightweight package.

What we learned

Verification must be deterministic—relying on LLM self-evaluation leads to false positives. Tracking source provenance (Primary vs. Secondary) at ingestion is critical for reliable intelligence.

What's next for NEXUS — Autonomous Workflow Intelligence

Adversarial Verification: A "Prosecutor Agent" tasked with finding counter-evidence before signing off on claims. Native Obsidian Plugin: Running NEXUS workflows directly within the Obsidian desktop app. Multimodal Chart Extraction: Extracting hardware block diagrams and financial graphs from PDFs.

Built With

Share this project:

Updates

Submission history