I build systems that turn complexity into clarity.

01 / Applied AI + research tooling

Researches 2–8 similar products, separates verified facts from vendor claims and bounded inference, then scores the shortlist against the way you actually work.

• Reports fit, confidence, and coverage as three distinct signals, so a closed-source tool can rank well on fit while staying explicitly low-confidence.
• Model output is constrained to a strict schema; weighted fit, freshness, conflicts, and privacy risk are recalculated deterministically in local code.
• No accounts and no hosted database—reports live in IndexedDB, and source collection is SSRF-hardened with IP-range, redirect, content-type, and size checks.

02 / Rust + developer tooling

Ingests shell history and AI-agent transcripts into local SQLite, mines the multi-step routines you keep repeating by hand, and drafts the automation for them.

• Four pattern miners separate repeated workflows, execute-fail-fix loops, recurring intents, and near-duplicate prompts—each drafted as a script, alias, or agent skill.
• Closes the loop: every decision snapshots history, so gain can measure real adoption and evolve revises or retires automations that never stuck.
• Single Rust binary with a dependency-free HTTP dashboard and a ratatui browser; the only network call is one Claude invocation, and only on request.

03 / Go + native macOS

A shared Go core powers both a Cobra CLI and a native SwiftUI companion app for inspecting, reproducing, and repairing macOS development environments.

• Snapshots Homebrew, npm, pip, VS Code, and shell configuration into portable restore plans.
• Analyzes orphaned packages, conflicts, disk usage, and upgrade risk.
• Converts natural-language setup requests into allowlisted, dry-run-first installation plans with explicit confirmation.

04 / Systems + AI

Concurrent feed ingestion with worker pools, rate limiting, and backoff—paired with exact and embedding-based deduplication through OpenRouter and Pinecone.

• Parallel embedding and vector search cut end-to-end latency by roughly 5×.
• Prometheus metrics and structured zerolog stats expose throughput, errors, and stage latency.
• Cost-aware model routing, optional knowledge graph extraction, PostgreSQL persistence, and Docker deployment.

05 / Rust + systems

A Mac-focused Activity Monitor alternative that combines live system metrics, process attribution, and safe process controls in one terminal UI.

• Attributes background processes to LaunchAgents, LaunchDaemons, and macOS applications.
• Supports search, sorting, process grouping, force quit, and confirmed process-tree termination.
• Separates collection, diffing, platform integration, TUI rendering, and scripting into a clean Rust workspace.

06 / Visualization + AI

Transforms public GitHub repositories into interactive 3D cities, then uses Claude to make unfamiliar architecture easier to navigate.

• Maps directories to districts and files to buildings sized by lines of code and colored by type.
• Combines architecture graphs, complexity scoring, contributor activity, and issue matching in one workspace.
• Streams AI-generated analogies, stack explanations, and personalized code-reading paths over SSE.