AntiTheca- The Adversarial Vault That Thinks While You Sleep

Inspiration

Every engineer, founder, and researcher suffers from the Bookmark Graveyard: we save hundreds of Instagram reels, architecture slides, technical tweets, and Obsidian notes, but when making high-stakes decisions, we cannot recall them.

Worse, existing personal AI assistants are dangerously sycophantic. If you tell ChatGPT or Claude, "I'm thinking of breaking my solo MVP into 20 microservices with Kafka," they respond with: "Great idea! Here is a Kubernetes manifest..."—even if three weeks ago you wrote an Architecture Decision Record (ADR 028) swearing never to introduce distributed microservices.

We built AntiTheca (Anti + Theca, "The Adversarial Vault") to solve both problems: an autonomous cognitive memory system that consolidates your knowledge while you sleep and acts as an uncompromising, zero-sycophancy devil’s advocate during high-stakes work.

What it does

AntiTheca is an autonomous personal AI system built around unprompted overnight synthesis and adversarial verification:

1. Autonomous 3:00 AM Sleep Engine: AntiTheca features an autonomous background daemon that fires every night at 3:00 AM without human prompts. It traverses newly ingested multimodal notes, detects topological bridges across concepts, identifies contradictions, verifies open technical questions via Tavily, and prompts NVIDIA Nemotron 3 Ultra 550B on Nebius Token Factory to synthesize an executive morning briefing.
2. Zero-Sycophancy Adversarial Radar: AntiTheca refuses to be an agreeable yes-man. When you ingest new reels or propose ideas that contradict your established principles, AntiTheca flags the tension, cites your historical ADR precedents, and forces you to defend or concede.
3. Multimodal Reel & Slide Extraction: Powered by NVIDIA Nemotron 3.5 Super-VL, AntiTheca parses screenshots, architectural diagrams, carousels, and video keyframes directly into structured epistemic memory.
4. Iterative Multi-Query Tavily Deep Research Loop: Decomposes complex claims into 3 parallel orthogonal subqueries (empirical benchmarks, failure modes, consensus standards), deduplicating web sources and computing an empirical confidence score.
5. Native PARA Vault & Obsidian Interoperability: Organizes notes into the PARA taxonomy (Projects, Areas, Resources, Archive), auto-generating bidirectional [[wikilinks]], Meta/vault-map.md, and an interactive 2D graph view.
6. Temporal Commitment Radar & Cognitive Defrag: Surfaces buried deliverables and deadlines from messy notes and cleans stale or orphan concepts via 5-phase defragmentation.
7. Static Epistemic AST Code Drift Auditor: Automatically walks the Python Abstract Syntax Tree (ast.NodeVisitor) to enforce codebase compliance with documented ADRs (e.g. database isolation boundaries and secret token sovereignty).

How we built it

• Inference & Compute on Nebius Token Factory:
  • nvidia/Nemotron-3-Ultra-550b-a55b: Epistemic arbiter for overnight consolidation and dialectical synthesis.
  • nvidia/Nemotron-3.5-Super-VL: Multimodal vision extraction for slide carousels, infographics, and reel keyframes.
  • nvidia/Nemotron-Mini-4B-Instruct: Rapid token chunking and metadata classification.
• Evidentiary Grounding: Tavily AI Search via an asynchronous 3-query decomposition engine providing empirical citations and consensus scoring.
• Data Sovereignty: Dual-database architecture using SQLite in WAL mode (self_brain.db for first-party convictions/ADRs; world_brain.db for external web groundings).
• Frontend Architecture: Next.js 16 with Vanilla CSS glassmorphic tokens, featuring an ultra-dense roaming synaptic cortex hero canvas, Obsidian-style cybernetic graph view, and real-time security telemetry.
• Security Isolation: NVIDIA OpenShell and eBPF sandbox guard isolating web ingestion scrapers.

Challenges we ran into

1. Model Sycophancy: Instruction-tuned LLMs naturally lean toward politeness. Forcing Nemotron 3 Ultra 550B to rigorously cross-examine user proposals against past ADRs required structured contrastive prompt engineering and counter-evidence retrieval pipelines.
2. Epistemic Contamination: Early prototypes mixed internet search results with personal convictions. We eliminated contamination by strictly isolating self_brain.db from world_brain.db at the database and AST validation layers.
3. Multi-Query Synthesis Latency: Running sequential web verification requests stalled consolidation. We refactored the Tavily client into an asynchronous asyncio.gather parallel search loop with URL deduplication, cutting latency by 65%.

Accomplishments that we're proud of

• True Autonomous Cognition: AntiTheca runs its 3:00 AM sleep replay cycle entirely unattended, turning passive bookmarks into structured decisions without human prompts.
• Instant Adversarial Detection: Processing an architectural diagram slide from an Instagram reel and having Nemotron 3.5 Super-VL flag a conflict against ADR 028 in under 3 seconds.
• Dual-Database Epistemic Sovereignty: Building a personal second brain that respects strict physical boundaries between personal truths and third-party web claims.

What we learned

• Cognitive memory is not vector similarity search; it is epistemic tension resolution. Simply retrieving top-k embeddings fails when notes contradict each other. An effective second brain must highlight contradictions rather than blend them into polite consensus.
• NVIDIA's Nemotron 3 Ultra 550B excels at complex dialectical arbitration when provided rich, unhurried batch context during overnight synthesis.

What's next for AntiTheca

• Morning Audio Briefing: Generating a 2-minute personalized podcast summarizing overnight thoughts using ElevenLabs.
• 1-Click Browser Extension: Instant bookmarking for TikTok and Instagram reels into the ingestion loop.
• Edge Inference on NVIDIA RTX: Running local quantized Nemotron models directly on workstation RTX 4090/5090 GPUs.

Built With

• abstract-syntax-tree
• asyncio
• autonomous-agents
• docker
• ebpf
• fastapi
• force-directed-graph
• multimodal-ai
• nebius-token-factory
• next.js-16
• nvidia-nemotron-3-ultra-550b
• nvidia-nemotron-3.5-super-vl
• nvidia-nemotron-mini-4b
• nvidia-openshell
• obsidian
• python
• react-19
• scikit-learn
• sqlite-wal
• tavily-ai-search