Explored. Tested. Shared.
On AI systems design, data architecture, and the software frontier — insights from the engineers working at the edge.
Model Abliteration
Model abliteration: the process by which a model’s safety guardrails get bypassed, exposing knowledge it was designed to refuse. A walk through how easy it actually is, why prompt-engineering alone isn’t enough, and what threats expand as model access democratizes.
read article →Recent articles.
Stumbling onto GPU Characteristics through Experimentation
When CUDA is overkill: using OpenACC pragmas as a middle path between CPU-only sequential code and full CUDA rewrites. Real benchmarks on Jetson AGX Orin showing the trade-offs, plus six practical use cases for modernizing legacy C/C++ on edge devices without full CUDA reengineering.
read article →A stalker named “Eigen”
Encountering eigenvalues in college vs. now using LLMs as patient explainers. On the joy of learning.
Native vs. Translated Prompt Engineering across Indian Languages
Sarvam-M vs. LLaMA-3 vs. Mistral on Hindi, Tamil, Telugu. Measured latency, cost, token distribution.
From the archive.
Cut LLM spend ~5× with vector-based caching and routing.
Dijkstra on natural-language programming, applied to vibe coding.
Building a healthcare RAG pipeline from zero, honestly.
Alternatives to matrix multiplication for low-power edge GenAI.
A custom Metal device plugin for GPU slicing on k3s/k8s.
Natural-language SQL queries with GPT-3.5 against CockroachDB.
A skeptic’s read on the Devin announcement and AI hype cycles.
Working on something we should write about?
Tell us what you’re building. The best perspectives come from real engagements.