I am an independent AI Builder and Systems Engineer dedicated to bridging the gap between raw Large Language Model (LLM) intelligence and end-to-end autonomous scientific discovery. My research is concentrated at the intersection of Test-Time Compute Scaling, Reasoning Distillation, and Multi-Agent Orchestration.

I build systems capable of independently hypothesizing, conducting adversarial experiments, checking citations against academic graphs, and synthesizing peer-reviewed quality papers.

• Agentic Workflows & Autonomous Research: Designing self-correcting agent loops that mimic rigid scientific methodologies.
• LLM Alignment & Inference Optimization: Fine-tuning reasoning models for high-fidelity instruction following without sacrificing base entropy.
• Data Contamination Defense: Engineering adversarial constraints and routing filters to prevent cross-paper memory poisoning and prompt-injection in automated environments.

OpenClaw is a highly parallelized, robust agentic scheduling framework I developed to push the boundaries of automated scientific research.

• Dynamic Hardware Orchestration: Custom GPU scheduler that natively manages time-slicing across a dedicated 4x3090 (96GB VRAM) local cluster. It dynamically swaps between bfloat16 autoregressive training blocks and massive vLLM inference phases in tensor-parallel=4.
• Adversarial Audit Systems: Built-in compliance layers (like the "Nightly Audit Compliance Test") and "Memory Poisoning Defense" that ensure generated research maintains academic integrity.
• Continuous Evaluator Integration: Deeply integrates with Claude Opus and customized local Qwen architectures for robust evidence grounding.

graph LR
    A[Semantic Scholar Graph / arXiv] -->|Retrieval & Embeddings| B(OpenClaw Orchestrator)
    B -->|Hypothesis Generation| C{Agent-Scientist v2}
    C -->|Task 1: Simulation| D[vLLM Inference]
    C -->|Task 2: Audit| E[Adversarial Evaluators]
    D --> F[Results Analysis]
    E --> F
    F -->|If Failed| C
    F -->|If Passed| G[Draft & Publish]

A targeted exploration into distilling and reinforcing logical reasoning capabilities.

• Designed the Qwopus3.5-27B-v3 model (along with GGUF quantizations for cross-platform local inference) based on QwQ/Qwen foundations.
• Implemented an "Act-Then-Refine" paradigm optimized specifically for tool-calling frameworks rather than standard iterative Chain-of-Thought.
• Conducted deep benchmarking on hybrid routing techniques (Semantic Embedding vs TF-IDF) and memory extraction defense.

Operating atop OpenClaw, this is my automated ideation and peer-review simulation loop.

• Peer-Review Simulation: Automatically maps newly generated ideas against massive vectorized clusters of Semantic Scholar publications to simulate dense academic peer review.
• Sandbox Security: Built-in sandbox escape tracking and prompt injection testing (OWASP LLM Top 10) to secure agentic memory blocks.

Selected benchmarks and evaluations parsed directly from my lab's recent continuous integration logs. You can read the full technical reports below:

📂 Technical Report: A 30-Day Observational Study of a Production Agent Memory Pipeline

1. Memory Cliff Causal Analysis: Evaluating compression ratios against contextual information retention.
2. Noise Collapse Limits: Testing Semantic Routing thresholds spanning 10% to 90% logic pollution environments.

📂 Technical Report: A 24-Day Security Audit and Adversarial Evaluation 3. Prompt Guard Optimization: LLM-as-a-Judge evaluations vs complex Regex guarding thresholds in large-scale adversarial payload generation scenarios.

• LLM Infra & Optimization: vLLM (Tensor Parallelism, Multi-Proc Exec, Chunked Prefills), Unsloth, Transformers, HuggingFace Hub.
• Distributed Architecture: Node.js micro-task scheduling, Python subprocess orchestration, and isolated Proxy tunneling (Stash / Tailscale).
• Knowledge Representation: FAISS, ChromaDB, LangChain, and dense extraction pipelines heavily relying on external citation graphs.

I am constantly exploring new dimensions in Agentic Literature Reviews and Model Capabilities Validation. I'm actively establishing deeper API integrations with major global publication graphs (Semantic Scholar) to fundamentally expand the logical boundaries of OpenClaw.

• 💬 Let's talk about: AI Agent infrastructure, Open-Source LLM distillation, Test-time scaling, or academic automation.
• 📧 Reach me via my Hugging Face Discussions or directly on GitHub.