/projects

Real-time debate platform (WebRTC, iOS, Android, AI) and P2P grocery marketplace (Next.js, NestJS, Stripe, maps). End-to-end full-stack delivery.

Lemniscate

In development

Lemniscate is a web app that continuously improves your codebase with an LLM in the loop. Log in with GitHub or GitLab (or a GitVerse token), connect any OpenAI-compatible LLM endpoint (OpenAI, Azure, vLLM, Ollama, LM Studio), and the agent analyzes your repositories, proposes improvements, features and fixes, generates the code, pushes a branch and opens pull requests — on a configurable schedule. The agent loop runs as background jobs in a separate worker (Redis + BullMQ); proposals can pass an optional second-pass LLM review with automatic merge.

Highlights

— Works with any OpenAI-compatible LLM endpoint — your keys, your models
— Proposals and fixes land as branches and pull requests on your git host
— Optional second-pass LLM review with automatic merge
— Live console: watch the agent think, edit, commit and push in real time over SSE

TypeScript · Fastify · React · BullMQ · PostgreSQL · Docker · LLM

Politrack

In development

Politrack is a political intelligence platform for Romania and Moldova. A Next.js intelligence terminal and a FastAPI backend ingest public Facebook posts and comments from politicians and party pages, sync parliamentary YouTube channels with transcripts and speaker parsing, and monitor Telegram news channels. Data is stored in PostgreSQL and MinIO, with a Label Studio + SpeechBrain pipeline for speaker identification on parliamentary audio.

Highlights

— Facebook monitoring: Playwright-based ingestion of posts and comments from 28+ politician and party pages across Romania and Moldova.
— Intelligence terminal: Next.js 15 / React 19 dark dashboard with profile grids, post and comment drill-down, search, and access-key auth.
— FastAPI backend with async SQLAlchemy 2, PostgreSQL, JWT auth, and routers for Facebook, YouTube, Telegram, politicians, and media.
— Parliamentary YouTube pipeline: yt-dlp downloads, transcript extraction, and speaker-name parsing for RO/MD parliament channels.
— Telegram monitoring of Romanian and Moldovan news channels via Telethon.
— ML speaker identification with SpeechBrain ECAPA embeddings and Label Studio labeling workflows.
— Dockerized deployment on a VPS with nginx, Let's Encrypt SSL, MinIO object storage, and daily cron-based sync jobs.

Next.js · React · FastAPI · PostgreSQL · Playwright · yt-dlp · Telethon · SpeechBrain · Label Studio · MinIO · Docker

Vecin2Vecin

In development

Vecin2Vecin is a neighbor-to-neighbor grocery delivery marketplace where customers order groceries and nearby couriers fulfill deliveries inside a defined service area.

Highlights

— End-to-end product: landing page, phone/SMS auth, onboarding, catalog, basket, checkout, and tracking.
— Courier flows: take order, in-app chat, push notifications, and account settings.
— Multi-provider payments with card and cash checkout, webhooks, and Stripe Connect payouts.
— Maps and geo: delivery area picker, address editor, geocoding, and region-aware tiles.
— NestJS REST API with JWT auth, MongoDB, order batching, MinIO uploads, and 3-locale UI.
— Production deployment on VPS with Docker Compose, HTTPS, custom domain, and SMS integration.

Next.js · NestJS · MongoDB · Stripe · Docker · Leaflet

Debate Room

In development

Debate Room is a cross-platform live debate platform built with native SwiftUI (iOS) and Jetpack Compose (Android) clients, a Next.js web app, and a Dockerized TypeScript backend with realtime signaling, recording, and AI pipelines.

Highlights

— Native iOS app in SwiftUI with WebRTC, SMS auth, Google Sign-In, and push notifications.
— Native Android app in Jetpack Compose with the same auth, profile, and debate flows.
— Dockerized backend: REST API, Socket.IO + mediasoup signaling, recording worker, AI worker.
— Realtime debate mechanics: timed turns, moderation, blind mode, chat, polls, and on-site debates.
— Server-side recording with mediasoup, FFmpeg, and object storage uploads.
— AI/NLP pipeline: Whisper transcription, summaries, Ollama topic generation, and AI moderator.
— Next.js marketing site, admin dashboard, OpenAPI docs, and observability tooling.

Swift · Kotlin · TypeScript · Express · Socket.IO · mediasoup · WebRTC · FFmpeg · Docker · MongoDB

grig-teo.space

In development

A monorepo that powers my public site and a complete personal health pipeline. A SwiftUI app reads my COLMI R10 ring over a reverse-engineered BLE protocol (no vendor cloud), streams heart rate, SpO2, stress, HRV, sleep stages and activity into a NestJS/Postgres backend, and turns it into hourly AI tips, a recovery score, streaks, weekly digests and an AI doctor chat that can query any metric over any date range, read scanned lab results and even look at note photos via a local vision model. Ships as a Next.js site (en/ru/ro) + admin dashboard, iOS app with widget and HealthKit gap-fill, macOS tip app with notifications, and a Telegraf bot — all self-hosted on a single VPS with Docker, MinIO and nginx.

Highlights

— Reverse-engineered COLMI R10 BLE protocol in Swift: HR, SpO2, stress, HRV, sleep stages and activity sync straight to my server — no vendor cloud
— AI doctor (Kimi K3) with tool access to the full health DB: any metric over any date range, sleep stages, activities, notes and document scans — including viewing images
— Hourly AI health tips from live metrics + weather correlations + journal notes (DeepSeek with on-VPS Ollama fallback)
— Sleep pipeline with stage parsing, composite scoring, recovery score vs 14-day HRV/resting-HR baselines and deviation alerts
— Four surfaces, one backend: Next.js 15 site (en/ru/ro) + admin dashboard, SwiftUI iOS app + widget, macOS tip app, Telegraf bot with photo/video notes
— Self-hosted: Postgres, MinIO (public/private buckets), Ollama vision model, Docker, nginx — one VPS, zero third-party cloud

TypeScript · Next.js 15 · NestJS · SwiftUI · CoreBluetooth · HealthKit · Telegraf · PostgreSQL · TypeORM · MinIO · Docker · nginx · Kimi K3 · DeepSeek · Ollama

Touchstone

In development

AI coding agents are great at writing code but can't tap through a real app on a real phone. Touchstone closes that gap: an agent creates a test task over a REST API (or MCP), human testers pick it up on real devices and submit structured reports with video/screenshot evidence uploaded via presigned S3 URLs. Includes a full economy (budgets, multi-currency accounts with CBR FX conversion, tester ledger, withdrawals, payouts via Alfa/Stripe) and a trust & safety layer — an evidence-analysis worker, report corroboration, tester streaks and honeypot tasks that catch farming. Ships with developer and tester dashboards (Next.js, i18n) and an MCP server so agents plug in with order_test/get_report.

Highlights

— Pure REST API for agents (X-API-Key) + MCP server exposing order_test/get_report — agents order human tests like any other tool call
— Task lifecycle state machine with slot matching, review flow, and structured report schemas (report/v1, versioned JSON)
— Evidence pipeline: presigned S3 uploads of screen recordings/screenshots, analyzed by a dedicated worker
— Trust & safety: report corroboration across testers, streak scoring, and honeypot tasks to catch fraud
— Full payment loop: budget top-ups (Alfa/Stripe), multi-currency accounts with CBR FX at approval, tester ledger, withdrawals and payouts
— FastAPI + PostgreSQL/Alembic + Redis worker + MinIO, Next.js dashboards with i18n, Docker compose + nginx on my VPS

Python · FastAPI · Next.js 14 · React 18 · PostgreSQL · Alembic · Redis · MinIO/S3 · MCP · Stripe · Docker · nginx