PeakTranslate Live
AI-powered video and document translation platform — peaktranslate.com
PeakTranslate is a subscription-based SaaS that lets users upload videos or documents and receive translated subtitles, AI-generated voice-over dubbing, and translated documents. It handles the full pipeline from media ingestion to delivery.
Architecture
Tech Stack
Key Technical Decisions
- Celery + Redis for async jobs — Translation and dubbing are long-running (minutes). Celery workers process jobs in the background while the API returns immediately with a job ID for polling.
- Alembic migrations — 15+ migration versions tracking schema evolution: users, subscriptions, credit system, billing events, voice-over modes.
- Stripe subscription billing — Three-tier plan system (Explorer, Pro, Business) with credit-based usage. Webhook-driven lifecycle management for subscription events.
- Multi-format pipeline — Single engine handles video (SRT/VTT subtitle extraction), audio (transcription), and documents (DOCX/XLSX parsing) through a unified job interface.
Backend Highlights
- Async SQLAlchemy sessions with connection pooling
- JWT authentication with Argon2 hashing + Google OAuth2
- OTP-based email verification via Resend
- Structured logging with rotating file handlers
- Pydantic request/response validation
- Cost estimation engine for pre-job pricing
Database Schema
Core models: User, Job, UserSubscription, SubscriptionPlan, BillingEvent, CreditTransaction. Normalized with proper relationships and constraints.
Deployment
PHOENIX Quant AI Live
Stock signal intelligence terminal — stocksphoenix.tech
An automated system that monitors Twitter/X and Reddit for stock ticker mentions, scores sentiment, generates trading signals, manages a simulated $100K portfolio, and produces AI-written equity analyst reports — all served through a real-time dashboard.
Architecture
Tech Stack
Signal Generation Pipeline
- Collection — Scrapes configured Twitter accounts and Reddit subreddits every minute. Extracts
$TICKERmentions with deduplication. - Filtering — Validates tickers against a 400+ term blocklist, then checks market cap (>$500M), price (>$2), and volume (>500K) via yfinance.
- Scoring — Aggregates mention frequency, source diversity, and sentiment into a composite score using logarithmic weighting.
- Signal — BUY when score exceeds threshold with sufficient source diversity. SELL when sentiment turns negative on held positions.
- Paper Trading — Enters positions with 2% of virtual capital. Exits on 10% take-profit, 5% stop-loss, or 5-day time limit.
Dashboard Features
- Live Screener — Real-time ranked ticker table with mention counts, sentiment, source breakdown
- Alpha Analytics — Four analysis tabs: Dormant Volcano (sudden activity spikes), Sentiment Divergence, Capitulation Detection, Source Grader
- Paper Trades — Open positions, closed trade history, portfolio P&L tracking
- Security Deep-Dive — AI-generated analyst reports with candlestick chart overlay
AI Reports
On-demand equity analysis via OpenRouter LLM. The synthesizer concurrently fetches financials, news, insider trades, and relative performance, then builds a structured prompt for the LLM to produce a full analyst report covering business profile, catalysts, risks, and capital decision.
Deployment
Nihongo.page Live
Japanese vocabulary learning platform — nihongo.page
A personalized Japanese language learning app built around spaced repetition. Users study 10 vocabulary words per day through interactive flashcards, get AI feedback on practice sentences, and listen to weekly generated audio scripts tailored to their vocabulary level.
Architecture
Tech Stack
Spaced Repetition System
Custom SM-2 variant with three learning phases:
- LEARNING — Initial exposure. Cards repeat at 1-minute and 10-minute intervals. Graduate to REVIEW after two correct answers.
- REVIEW — Long-term retention. Interval grows based on ease factor (minimum 1.3). Rating scale: Again (reset), Hard (×1.2), Good (×ease), Easy (×ease×1.3).
- RELEARNING — Failed review cards re-enter short intervals before returning to the review queue.
The algorithm tracks interval, easeFactor, repetitions, dueDate, and nextShowAt per user-card pair. Cards rated "Again" or "Hard" are re-inserted into the current session queue immediately.
Weekly Audio Scripts
Each week, the system generates 3 audio scripts using the user's learned vocabulary:
- Script 1: 60% current week's words + 40% new vocabulary
- Script 2: 80% full library + 20% new
- Script 3: 100% from existing library (reinforcement)
Users choose script types: Story, Dialogue, News Broadcast, Travel Scene, or Action Adventure. Text is chunked by Japanese punctuation (max 300 chars), sent to Qwen3-TTS in parallel, concatenated, and stored on DO Spaces.
Database Schema
10 Prisma models: User, Card (2,000+ entries from Kiashi deck), UserCard (SRS state per user), DailySession, UserSentence (with AI feedback), WeeklyScript (with audio URLs). Part-of-speech enum: Noun, Verb, Adjective, Adverb, Particle, Other.
Features
- Interactive flashcard study UI with progress bar
- AI-powered sentence writing practice with grammar feedback
- Vocabulary library browser with kanji radicals deck
- Dashboard with learning analytics
- Dark mode, onboarding flow, Hiragana keyboard helper
- Self-hosted Umami analytics at
analytics.nihongo.page