╔══════════════════════════════════════════════════════════════╗
║           building the infrastructure layer for agents        ║
╚══════════════════════════════════════════════════════════════╝

 agents need boundaries.    queries need answers.    markets need clarity.
       |                           |                        |
    Kernex                       Mercer                  Cynosure

"Don't trust the model. Trust the kernel."

OS-level execution sandbox for AI agents. A single statically-compiled Rust binary that intercepts syscalls in real time using Linux Landlock LSM + seccomp BPF — before the agent process even boots. No VMs. No daemons. No code changes required.

# Before
python my_agent.py

# After — fully sandboxed at the OS level
kernex run -- python my_agent.py

• Audit mode — auto-generates a least-privilege kernex.yaml policy by observing one run
• < 2ms boot overhead vs ~500ms for Docker
• MCP co-sandboxing — each MCP server gets its own independent policy
• JIT interception — blocked actions prompt the user rather than crash the agent

Rust · Landlock LSM · seccomp BPF · macOS Endpoint Security · Unix Domain Socket IPC

Plain English to accurate SQL, even on schemas with cryptic abbreviations (cust_seg_cd, e_add, p_spec), missing foreign keys, and inconsistent naming. Six-stage agentic pipeline running entirely on a consumer GPU — no vector database required.

Question → [Entity Retrieval] → [Schema Linking] → [Query Decomposition]
        → [Candidate Generation x3] → [Execution + Scoring] → [Correction] → SQL

Benchmark (Qwen2.5-Coder-7B, RTX 4070 Laptop):

Python · llama.cpp · Qwen2.5-Coder-7B · Triton · FastAPI · React · Redis · SQLAlchemy

"Follow the star. Trade with clarity."

Autonomous perpetual swap trading system for OKX — crypto majors, gold/silver, US equity index perps. Zero cloud LLM dependency. A local Qwen3.5-4B synthesizer reads a pre-computed expert signal brief (~500 tokens) and outputs a single structured JSON decision; all risk logic runs in deterministic Python.

15-min cycle:
  Expert Pipeline → MarketBrief (~500 tok) → LLM Synthesis (5-8s)
  → Signal Persistence Gates → Risk Engine (Kelly sizing) → OKX Execution

Signal sources per cycle: EMA/RSI/MACD/OFI across 3 timeframes · TimesFM 2.5 zero-shot forecast · L2 orderbook depth · Fear & Greed · funding rates · open interest · liquidation clusters

Python · Qwen3.5-4B · Ollama · TimesFM 2.5 · OKX MCP · SQLite · APScheduler

500M parameter lip-reading model. VALLR-based architecture that transcribes speech from video without audio.

Python · PyTorch · VALLR

systems programming  ·  LLM inference  ·  agentic infrastructure  ·  kernel security