Technologist.
Experimenter.
Seeker.

✓ Sovereign personal knowledge and automation system

A local-first intelligence layer built around an Obsidian-compatible Markdown vault. Combines durable AI agents, multimodal capture, layered lexical retrieval, evidence-backed actions, natural voice interaction, rebuildable indexes, and local Cogni-Brain inference on DGX Spark.

✓ High-context multimodal agent brain on DGX Spark

Qwen3.5-122B-A10B INT4+FP8 hybrid served via vLLM with DFlash speculative decoding, text/image/audio inputs, 262K local context, 54.44 tok/s single-stream (68 tok/s peak), and 100/100 Tool-Eval. The latest high-context Cogni-Brain iteration for local agentic and multimodal workflows.

✓ Terminal/action workbench for Cogni-Brain

A self-correcting bash/action agent using a Qwen2.5-Coder-14B model fine-tuned on NVIDIA Nemotron SFT data. Explores ReAct-style command execution, terminal feedback loops, workspace-scoped runs, output limits, and safer tool-use guardrails for local agent systems.

✓ Sustained long-context agentic work on DGX Spark

Nemotron-3-Super-120B pushed beyond one-click demonstrations into sustained agentic workloads. Stabilised at 131K local context, approximately 24 tok/s, and 93/100 Tool-Eval, with a hardened agentic stack and reproducible benchmark methodology.

✓ Fast local tool-agent brain on DGX Spark

Qwen 3.6-35B served through Atlas with approximately 219 tok/s, 100/100 Tool-Eval, and local NVFP4 acceleration. Demonstrates how a smaller model paired with an efficient inference architecture can support responsive coding and agent loops.

✓ Dedicated local multimodal perception agent

Gemma 4 12B on DGX Spark via vLLM with image input, audio and voice-note workflows, video-as-frames perception, multilingual chat, tool calling, 196K tested context, and reproducible local benchmarks.

✓ Local-first multimodal AI ecosystem

A broader ecosystem 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.

✓ Sanskrit language model trained from scratch on DGX Spark

A 7.5M-parameter decoder-only Transformer trained on approximately 140M pure Sanskrit tokens, with 6 layers, 8 attention heads, 256-dimensional embeddings, a 1024-token context window, and an approximately 8K vocabulary.

✓ Sanskrit language model trained from scratch on Apple Silicon

A 7M-parameter decoder-only Transformer trained on approximately 57M Sanskrit tokens using Apple MLX, with 4 layers, 8 attention heads, 256-dimensional embeddings, and an 8K vocabulary.

✓ Autonomous enterprise governance agent

Implements a strict Check-then-Act protocol using Gemini 2.5 Flash Lite to enforce real-time fiduciary risk alignment through tool-first orchestration. Built as a capstone project for the Google and Kaggle AI Agents Intensive.

✓ Indian Sign Language recognition

Computer vision model achieving 99% precision and recall on Indian Sign Language. Training data was collected personally in 2018 to explore technology that could bridge communication for deaf and non-speaking communities.