technologist | experimenter | seeker

"the future belongs to those who understand at a very deep level how to combine their unique expertise with what algorithms do best" - Pedro Domingos, The Master Algorithm

Independent Research & Systems.

Building sovereign, local-first AI systems - from training small language models from scratch, through multimodal, high-speed, and long-context agent research on DGX Spark, to Cogni Life OS: a private intelligence and automation layer for personal and household knowledge.

🧠 The Research: Building a Sovereign AI Stack

Single NVIDIA DGX Spark Node (GB10 · 128 GB Unified Memory)
_{Tiered Range: 12B → 27B → 30B → 35B → 118B → 120B → 122B → 176B   |   🟢 Current Daily Driver  •  🟡 Previous Stable}

Model	Params	Serving Engine	Context	Throughput	Tool-Eval	Multimodal	Spark Arena Benchmark
🟢 Qwen3.8 Flash Next	~176B-A6B	SGLang + HashK R=4 + NEXTN	~262K	~37 tok/s	100/100	Y	Verified Run ↗
Qwen3.8	~27B	vLLM NVFP4 + MTP	~262K	~24 tok/s	100/100	Y	Verified Run ↗
Muse-Glimmer (Inferact)	~30B	vLLM W4A4 + DFlash	~131K	~28 tok/s	90/100	Y	Verified Run ↗
🟡 Qwen 3.5	~122B-A10B	vLLM hybrid INT4+FP8 + DFlash	~262K	~54 tok/s	100/100	Y	Verified Run ↗
Qwen 3.6	~35B-A3B	Atlas NVFP4	~131K	~219 tok/s	100/100	Y	Verified Run ↗
Nemotron-3-Super	~120B-A12B	vLLM NVFP4	~131K	~24 tok/s	93/100	Text only	Verified Run ↗
Nemotron-3.5-Lightning	~30B-A3B	vLLM DSpark-NVFP4	~262K / ~1M	~112 tok/s (262k) ~42 tok/s (1M)	80/100	Text only	Verified Run ↗
Laguna-S-2.1 (poolside)	~118B-A8B	vLLM DFlash-NVFP4 (7 tokens)	~262K	~27 tok/s	97/100	Text only	Verified Run ↗
Gemma 4	~12B	vLLM BF16 KV FP8	~196K	~22 tok/s	83/100	Y	Verified Run ↗

🛠️ Systems & Ecosystem

• Cogni Life OS: Sovereign, local-first personal knowledge and automation system built around an Obsidian-compatible Markdown vault, with durable AI agents, multimodal capture, layered lexical retrieval, evidence-backed actions, natural voice interaction, and local Cogni-Brain inference on DGX Spark.
• CogniBash: Experimental terminal/action workbench for Cogni-Brain, exploring local ReAct-style bash execution, command feedback loops, workspace-scoped runs, output limits, safer tool-use guardrails, and fine-tuned code-agent behaviour on DGX Spark.
• SageGPT-7.2M (DGX): Sanskrit-only decoder-only Transformer SLM trained from scratch on a 139 MB tokenized Sanskrit corpus containing 72.8M SentencePiece model-token IDs, derived from a 105.2M-character purified Sanskrit text corpus. Runs on NVIDIA DGX Spark with 6 layers, 8 attention heads, 256-dimensional embeddings, a 1024-token context window, and an ~8K vocabulary.
• SageGPT-7M (MLX): Small language model trained from scratch on ~57M Sanskrit tokens using Apple MLX, with 4 layers, 8 attention heads, 256-dimensional embeddings, and an 8K vocabulary for Apple Silicon inference.
• Cogni.chat: Broader local-first multimodal AI ecosystem direction for personal and household intelligence, spanning memory, planning, wellbeing, learning, creative work, family coordination, and interaction across text, voice, images, and personal context. Cogni Life OS provides its sovereign knowledge and automation foundation.
• Fiduciary-Ops-Agent: Autonomous enterprise governance agent using a strict Check-then-Act protocol via Gemini 2.5 Flash Lite; enforces real-time fiduciary risk alignment using tool-first orchestration. Also works with local Cogni-Brain inference on DGX Spark.

---

_{"I have no special talent. I am only passionately curious" - Albert Einstein}