Blog
Explaining the underlying mathematics, sharing raw benchmark results, and analyzing codebases. No summaries, no filters.
Parallax: How I Built a CI Tool That Reads Your SQL and Maps Every Downstream Dashboard Before Your PR Merges
A walk-through of how Parallax works internally: SQL AST diffing, in-memory DAG traversal, column-level lineage tracing, and risk-scored PR blocking for dbt. Zero database credentials required.
NeuroLambda: How I Ran Two AI Models on One 8GB GPU to Build a Real-Time SRE Pipeline
Two models. One 8GB GPU. 707 minutes of continuous operation. NeuroLambda combines Mamba S6 (0.41ms/event) and Qwen-3B for automated root-cause diagnosis, achieving F1=0.9713 on zero-shot microservices data with 51% lower latency than static routing.
Why I Replaced Transformers with Mamba S6 for Log Analysis (70× Faster, Flat Memory at Any Length)
GPT-2 crashes OOM past 8,192 tokens and processes only 35 logs/sec. Mamba S6 runs at 2,453 logs/sec with constant memory regardless of sequence length. We built two 125M SSMs, fixed a 3GB streaming KV cache bug, and benchmarked both against GPT-2 across three random seeds.
I Trained a 124M-Parameter LLM to Catch Silent System Failures: Zero Labeled Data
No labeled failures. No regex rules. A 124M-parameter GPT-2 variant trained only on normal system logs detects anomalies at F1=0.8923 on a consumer RTX 3060 Ti by measuring how surprised it is by each log sequence.
From Alert to Fix: Fine-Tuning a 3B LLM on a Consumer GPU to Generate Automated Root-Cause Diagnoses
A 3.09B parameter LLM fine-tuned in just 28 minutes on a consumer RTX 3060 Ti. It reads raw system log blocks and outputs structured JSON with severity level, root cause, and exact CLI remediation commands. No 24GB server required.