This is not an expert-only panel. Discovery marketplace and operations live on the same data model. A member is tied to one live expert at a time. Request approval is locked inside a transaction with SELECT FOR UPDATE. A second expert gets 409 MEMBER_HAS_ACTIVE_EXPERT. A library template is deep-cloned onto the member. Catalog exercises are shared. A package ends on date only. Running out of sessions does not expire it. The subscription row is updated in place on renewal, so accounting keeps a separate subscription_sales ledger. Program archive is not deleted. Restore is locked to expert_id.

No external migration tool. Numbered SQL at boot plus _fitsync_migrations. Each file is one transaction.

No Stripe or İyzico for member-expert money. Snapshot price plus subscription_payments, remaining via SUM. Default currency TRY.

Chat writes over REST. An open screen uses one-way WebSocket. Fan-out is an in-process Map (single-instance note in the code). Conversation UUID pair is unique with user_id_a < user_id_b.

Push is not fire-and-forget. A notifications outbox row, minute dispatch cron. Tests do not call FCM, Resend, or S3.

Discover does not look at platform subscription. Active onboarded experts with a published price. Rank: 0.6 conversion count + 0.4 conversion recency (365/180 days).

• Expert signup. Password reset via 6-digit email OTP (Resend). No refresh rotation. Login drops the old refresh
• Discover and requests. Plan request, accept/reject, synthetic renewal (hidden from the member, EXPERT_RENEWAL_IN_PROGRESS)
• Library. PT/DT composite ref/new, assign to member, measurements, feed, session calendar, ledger/debtors
• 1:1 chat. Text + image. Conversation stays open after the package ends. Expert starts it
• Deno.cron. Reminders, outbox, İstanbul 15:00 water reminder (12:00 UTC), orphan chat images, hourly package expire, 15 min session complete
• OpenAPI + Scalar /docs. OpenTelemetry. System metrics in production

Ordu Genç İş İnsanları Derneği needed a private member app, not a public social network. ~180 invite only members, six roles (root → üye), and board approvals that must survive presidency transfers without orphaning queues. Shares, sector purchase requests, and dues all needed different feeds, notification rules, and audit trails on a mobile only surface with unreliable push delivery and Turkish calendar billing.

Role based approvals, not person based queues. Post and purchase requests are approved by yk_baskan / yk_vekili roles so a presidency transfer does not break pending work.

Separate share vs purchase feeds. Event style posts and sector buying announcements never merge in the API. Different schemas, paths, and notification targets.

Deno + Yelix/Hono + MongoDB. OpenAPI first REST for a Flutter client, Zod validation end to end, Coolify self hosting, R2 for media, FCM with retry cron, İstanbul aware dues jobs.

Root Admin Ask with read only Mongo. OpenRouter chat can query association data only through a readonly URI, so ops Q&A cannot mutate production state.

• Member directory & companies. Invite only accounts, company profiles, sector tags
• Share feed. Moderated posts with media, comments, redflag filtering
• Purchase requests. Sector buying pipeline → separate announcement feed
• Dues (aidat). Monthly rows, role multipliers, partial payments, reminder cron
• Notifications. In app inbox + FCM device tokens, event level push prefs
• Governance. Soft archive for members, hard delete + audit for companies
• Admin Ask. Root only AI over read only aggregates

A small fine-tuned adapter on top of Kara Kumru (AlicanKiraz0/Kara-Kumru-v1.0-2B). Trained on a curated subset of an internal benchmark for tool routing: deciding when a query needs an external tool versus a direct model reply. Metrics below are after applying the adapter; the base weights stay unchanged.

Binary labels only: TOOL if an external tool is required, DIRECT if the model can answer without tools. Dataset language: Turkish (mixed domain). Each sample is decision-only — not tool name or arguments.

{
  "instruction": "User query here",
  "output": "TOOL"
}

Load the base checkpoint, apply this adapter for routing, then handle tool schemas and full fine-tuning elsewhere. Next step: extend the adapter so that after TOOL, the model also picks which tool to call.

A location-based chat application and information platform specifically designed for the Ordu province in Türkiye. The project provides local services including transportation, news, places search, and regional information through an AI-powered chat interface. Built as a full-stack application with separate backend and frontend components.

Backend (Deno + Hono) — API server with chat functionality and data crawling capabilities.

Frontend (Astro) — Web interface with mobile app capabilities through Capacitor.

Database — MongoDB with Mongoose ODM for flexible data storage.

AI Integration — OpenRouter for chat functionality with tool-based pipeline.

• AI Chat Interface — Core functionality with tool-based pipeline
• Local Transportation — Public transport routes (Kentkart integration)
• Places Search — Mapping and location services
• Local News — Aggregated regional news
• Inter-district Transport — OrduBirlik integration
• Parking Availability — Real-time parking information
• Tourist Attractions — Curated local points of interest
• Bus Routes — Comprehensive transportation information
• Pharmacy Services — On-duty pharmacy schedules
• Restaurant Recommendations — Curated local dining options

The system includes a comprehensive data pipeline with web crawling for local information, scheduled data updates, and MongoDB storage with various data models. This ensures real-time and accurate information for all local services.

Native Android and iOS applications through Capacitor with geolocation services and native app distribution. The mobile apps provide the full functionality of the web platform with optimized mobile experience.

A repository-aware maintainer assistant for GitHub projects. co-maintainer analyzes a repository's code, workflows, commits, pull requests, discussions, and diffs to generate contribution guidance and surface likely review issues before human review.

Repository-specific contribution rules, review expectations, testing requirements, and workflows are difficult to discover and apply consistently. Generic coding rules miss the conventions that make a particular codebase maintainable.

Repository-aware analysis — guidance is derived from the project's own code, workflows, history, pull requests, and review discussions.

Low/high model separation — extraction and synthesis use different model paths so repository analysis can stay focused and composable.

Incremental cache — SQLite tracks codebase and PR revisions, avoiding unnecessary reprocessing while keeping later reviews current.

The CLI collects repository context, extracts patterns and expectations, synthesizes a repository-specific SKILL.md, and can use that guidance to review an individual pull request. Optional review guides, improvement passes, and detailed token, cost, timing, heartbeat, and debug logs make the process inspectable.

A second-pass review used approximately 38.9K input tokens, 11.4K output tokens, and cost $0.0234.

The CLI has reached a working stage and produces actionable pre-review findings. The main remaining evaluation area is measuring finding accuracy with feedback from engineers familiar with the target codebase.

A live terminal tracer for Express 5. Run a collector, attach a reporter, and watch requests, middleware timelines, and logs in a three-pane TUI. The instrumentation layer lives separately in express-emits, so measurement and presentation stay independent.

Express middleware is often a black box during local development. When a request is slow or incorrect, it is difficult to see which middleware ran, in what order, how long each one took, or which log line belongs to that request. Full APM and OpenTelemetry setups are too heavy for this kind of feedback loop.

express-emits instruments the app by wrapping app.use and emitting typed start/end events with real durations. express-trace-cli consumes those events, batches them over HTTP to a local collector, and renders sessions, requests, middleware trees, and logs in Ink. If the CLI is not running, events are dropped: this is a development tool, not a durable pipeline.

Request timing ends on res finish or close, not when next() returns. Middleware duration is self-time until handoff, not the downstream chain. Every middleware.start has exactly one end event, including sync throws, rejected async middleware, and early responses.

Five local autocannon scenarios compared tracing against vanilla Express. Counter-based IDs with zero listeners stayed around 16.5–16.9k req/s against a roughly 17–18k baseline; silent listeners and UUID IDs added somewhat more overhead.

A frame-accurate Windows video player built for VFX review. Drop in a file or the portable executable, paste a URL, step through frames, inspect the waveform, set in/out points, loop, trim, and compare A/B versions without opening a full NLE.

Consumer players are not designed for precise VFX review. Their playheads are difficult to trust, single-frame stepping is inconsistent, and local checks often turn into a codec, installer, or GPU-tool hunt. VFX Player keeps the review loop in one small Windows tool.

Frame stepping — move by ±1 or ±10 frames while seeing seconds, milliseconds, source FPS, and playback FPS to two decimals.

Review controls — zoom to pixels, waveform, volume from 0–125%, in/out loop, trim in the source format, and A/B wipe or side-by-side comparison.

Ingest and updates — FFmpeg and yt-dlp are bundled into the executable; the in-app updater replaces the binary from GitHub Releases.

FFmpeg and yt-dlp are compressed into the binary with zstd and extracted on first run. Video2X remains optional: it downloads once after consent, lists available GPU devices, and opens the result as the B layer. The current release is v0.1.8.

A free, no-setup review player used as a daily tool: open, step, mark, loop, and send.