Skip to content
~/writeups

Deep-dives with status, metrics, and update logs.

Civic Lens — System Architecture

shippedJul 9, 2026

How Civic Lens turns raw news, Reddit, and X into audited discourse metrics — a crash-safe Go crawler, a content-addressed blob store, one SQLite database shared across two languages, a five-engine Python analysis pipeline, and a snapshot-cached API.

GoPythonSQLiteFastAPIReactArchitecture

Civic Lens — Database Schema & Migrations

shippedJul 9, 2026

One SQLite database shared by a Go crawler and a Python analysis pipeline — 16 tables across ingestion, corpus, analysis, and ops; 21 transactional migrations; and foreign keys enforced on both sides so traceability is a constraint, not a convention.

SQLiteGoPythonData ModelingMigrations

Civic Lens — System Invariants

shippedJul 9, 2026

The non-negotiable correctness contract behind Civic Lens — traceability from every pixel back to raw bytes, evidence spans that must be verbatim quotes, confidence on every AI output, and proxy metrics that are never allowed to pose as ground truth.

CorrectnessDesignData IntegrityAI Safety

Civic Lens — LLM Evals: Golden Set + CI Regression Gate

in progressJul 9, 2026

How Civic Lens verifies that a prompt, schema, model, or pipeline change didn't silently regress LLM output quality — deterministic span-overlap scoring against a human-verified golden set, replayed recordings in CI, and prompt fingerprinting so stale recordings can't pass.

LLMEvalsCI/CDPythonTesting

Civic Lens — Scoring Methodology

shippedJul 9, 2026

How every number on the dashboard is computed and bounded — blended bot scores with human-readable indicators, lexicon + LLM sentiment, GOP favorability as an explicitly-labeled proxy, and propaganda detection that refuses to flag anything it can't quote.

LLMNLPPythonScoringData

Crawler + Sentiment Pipeline

pausedJun 30, 2026

A polite, crash-resumable web crawler in Go feeding a hybrid sentiment-analysis pipeline in Python, deterministic signals refined by a pluggable LLM, every output traceable to verbatim source text and gated by a human-review loop.

GoPythonWeb CrawlerNLPPipelines

Mini Inference Engine

in progressJun 30, 2026

A from-scratch LLM inference serving core in C++20/CUDA including a paged KV-cache allocator, a continuous-batching scheduler, and paged attention. Built to implement the load-bearing parts of a modern engine rather than import them.

C++CUDASystemsLLMInference

This Site (kobeyoung.net)

shippedJun 30, 2026

A portfolio, technical blog, and live self-hosted LLM demo built as a monorepo with a Next.js front end, a zero-dependency Go backend, and layered abuse defenses, self-hosted on a single Hetzner box.

Next.jsGoMDXSelf-hostedSecurity
Writeups — Kobe Young