SneakStream is a coordinated livestreaming platform for nightclub and DJ events, built around the bet that live music venues need broadcast-style synchronized playback, pay-per-view monetization, and a social layer — not just a generic video host.

The system spans a Convex-backed web client, a React Native mobile app (iOS and Android), and an admin layer covering broadcast playlist control, live auction bidding on track crates, VIP guest lists, Stripe and Apple IAP payment verification, and Mux webhook-driven recording ingest with full deduplication. A party photo capture loop — shoot, live rail, gallery wall, QR landing page, with HEIC transcoding for iPhone uploads — was added to extend the platform to event photography. Recent work has been almost entirely iOS-native: a full Dynamic Island Live Activity covering all six states, automatic brownout recovery, always-on activity startup, and frame-accurate replay-position sync between the player and the island widget.

compact is an experiment harness built around the problem of accurately capturing and grading discrete units of shipped work, with the bet that clean deduplication and multi-granularity aggregation can surface a reliable signal from noisy, redundant source data.

The core system ingests work entries at multiple granularities (commit, PR, weekly) and grades them; the recent focus has been a post-hoc merge pass that collapses duplicate beat descriptions across source units before re-grading survivors. Shipping that deduplication layer revealed a 34% duplicate rate at commit granularity — confirming the problem was real — while PR and weekly granularities were already largely clean at 7% and 0% respectively.

SwingAnalyzer bets that vision-capable LLMs can reliably evaluate athletic swing mechanics—the project exists to benchmark which models are actually useful for that task and to build the evaluation infrastructure needed to find out.

The core shipped work is a multi-model evaluation harness that runs a two-pass locate-then-refine vision pipeline across any OpenRouter-hosted model, generates a 3×3 contact sheet per model/video pair, and feeds those sheets to an independent LLM judge that scores each one 0–10. Results are collected into a markdown leaderboard and a structured results.json in a single run, making model comparison repeatable and auditable.

OpenParent is a household behavior-management system that uses an AI judge to score observed interactions between parents and children, automatically applying or flagging consequences against a configurable rubric and policy — betting that transparent, rule-governed feedback loops are more consistent than ad-hoc parental judgment.

The core stack is a TypeScript/Convex backend with a Python edge layer for audio and LLM work: the judge ingests transcribed kitchen exchanges, scores them against a versioned YAML rubric, and routes every consequence through a deterministic safety rail enforcing magnitude clamps, rate limits, and daily caps before anything touches the ledger. Recent work closed the full parent review loop — a pending-proposals inbox, per-event drill-down with Apply/Skip actions, a browser-based simulate console for rubric testing without hardware, and an on-device speaker-verification pipeline using cosine similarity on Resemblyzer embeddings. The test harness now covers the consequence rail, ledger idempotency, and proposal queries entirely in-memory with no live secrets required.

vibe-debug bets that coding agents produce better fixes when given real runtime state rather than static source inference — the project exists to wire a full DAP-backed Python debugger into the MCP tool-call layer so agents can observe live execution instead of guessing.

The core deliverable is an MCP server that exposes debugger operations (launch, set breakpoints, step, inspect locals) as structured tool calls, backed by the Debug Adapter Protocol. On top of that, an npx shim removes the Python toolchain prerequisite for agent registration, a stream-JSON formatter renders Claude debugging sessions as human-readable output, and a standalone CLI mode with an auto-written skill file lets agents discover and invoke the debugger without MCP at all. The consistent focus is reducing friction at every integration point — installation, invocation, and observability.

AllyBi is a video-statement platform built for both web and mobile, aiming to let authenticated users record and retrieve personal video content tied to a verified identity.

The web app received a complete Clerk-backed auth layer — eight screens covering sign-in, sign-up, MFA, SSO, and invite flows — with server-side userId stamping on uploads and an anonymous upload fallback. The mobile app followed with its own Clerk FAPI integration: secure-store JWTs, route-level auth gating, Google and Apple OAuth with browser-session handling, an OAuth phone-collection screen for incomplete sign-ups, and an SMS OTP second-factor screen to close out the MFA flow. Recent work has been focused on closing every dead-end in the mobile auth path and converging both platforms on a shared authenticated video feed gated by real Clerk JWTs.

WineMail is a personal wine management tool that treats the inbox as a primary data source — parsing merchant emails, order PDFs, and bottle photos to keep a cellar inventory, tasting diary, and live shop search in one place.

The core build covers merchant email ingestion with LLM-backed offer analysis, PDF order parsing into structured inventory records, AI image generation for wine profiles via FAL flux models, and a daily wine recommendation engine with food-pairing and label-photo input. Recent work has pushed outward: an MCP server exposes inventory data to external AI assistants (Claude Desktop, ChatGPT), dual-protocol support adds OpenAI Apps SDK widget rendering inside conversations, and a batch image generation pipeline with an admin control panel brings AI bottle art to the full inventory automatically.