v0.14.0 - Mobile-First Redesign & Agent Reliability Fixes

• Site-wide mobile responsiveness overhaul: sidebar becomes a slide-in hamburger drawer below the md breakpoint (was a static 208px column that ate over half the screen on a phone), chat widget becomes a full-width bottom sheet on phones instead of an off-fitting floating panel, and page padding across every route switched from a fixed 40px gutter to a responsive clamp()
• Fixed silent truncation across 3 LLM agents (Critic Review, Portfolio Construction, News Sentiment) where max_tokens was too small to complete a multi-part response - "Missing Risk Factors" and "Key Risks" sections were rendering empty; also fixed a parsing bug that dropped every line of a summary after the first
• Fixed a real regression in the multi-symbol parallel benchmark: os.cpu_count() reported the host machine's total cores instead of the container's actual cgroup CPU quota, causing the "parallel" demo to run slower than sequential on a quota-limited deployment instead of faster
• Fixed an N+1 query pattern in the price-history endpoint (one DB round-trip per row during a full historical re-ingest) and pushed date-range filtering into SQL - a full history fetch that could take minutes now completes in ~3 seconds
• New "How It Works" page: an interactive, expandable explainer covering all 9 major platform subsystems (signals, regime detection, ML forecasting, risk/pricing, portfolio construction, the agentic research pipeline, backtesting, the C++ layer, real-time streaming) in plain English, with an animated data-flow diagram
• Chat widget redesigned: dropped drag-anywhere positioning in favor of a fixed bottom-right corner everywhere, with a clearer "I'm an AI-powered assistant" greeting
• Bulk ticker-universe ingestion endpoint plus a scheduled GitHub Actions workflow to keep price history current automatically, instead of relying on lazy per-request ingestion
• Loading screen's minimum hold tuned to 30-35s (down from 40-45s) so the fun fact / mini-game / portfolio link are genuinely visible without overstaying
• 222 tests passing

v0.13.0 - C++ Acceleration Layer Improvements

• Extended C++20 kernel suite: added rolling_max, rolling_min via monotonic deque (O(n)), plus ATR kernel with Wilder smoothing
• Expanded numerical equivalence suite with per-kernel tolerance bounds verified to ≤ 4×10⁻⁹ vs pandas baseline
• Benchmark suite extended with multi-symbol parallel measurements (ThreadPoolExecutor) with 39.8× end-to-end on 8 × 1M rows
• CMake/scikit-build-core packaging improvements with cleaner pip install ./backend/cpp experience
• Graceful degradation confirmed: falls back to pandas silently when .so not built (Render production path)
• 152 tests passing

v0.12.0 - AI Chat Widget

• Floating chat button on every page
• Powered by Groq (free tier, llama-3.3-70b-versatile) with tool use
• 10 tools exposed: signals, regime, ML forecast, VaR, thesis agent, news sentiment, earnings, portfolio construction, backtest, factor model
• Tool badges show which algorithms were called for each answer
• Suggestion pills for quick onboarding
• Answers grounded in real AEQUITAS data — never hallucinated numbers

v0.11.0 - Animated Loading Screen + Infrastructure Migration

• Animated loading screen with typewriter finance facts, catch-the-ticker mini game, portfolio link, skip button, and auto-dismiss on API connect (90s hard cap for Render free-tier cold start)
• Infrastructure migrated to $0/month: Railway → Render + Neon (serverless Postgres) + Upstash (Redis)
• Keep-warm GitHub Actions workflow prevents Render idle spin-down
• Infrastructure-agnostic Alembic migrations (TimescaleDB hypertable applied only when extension is available)
• 152 tests passing

v0.10.0 - Three Agentic Features + Earnings Data Handling

• News Sentiment agent (Finnhub headlines + Groq LLM scoring, trend detection)
• Earnings Analysis agent (Finnhub calendar/fundamentals + Groq LLM synthesis)
• Portfolio Construction agent (mean-variance optimisation + cointegration + LLM thesis)
• New combined Agents page with 3 tabs
• Fixed: earnings stat cards now hide cleanly instead of showing placeholder "No data" when Finnhub's free tier has no calendar data for a ticker
• 152 tests passing

v0.9.2 - Factors UI, WebSocket Fix, Real-Time Polish

• Fixed WebSocket subscription race condition (early subscribe() calls were silently dropped if socket hadn't finished connecting)
• New Factors page -> Fama-French 3-factor model and TWAP/VWAP/Implementation Shortfall execution algorithms now have full UI, previously only accessible via /docs
• Execution schedule visualised as share-distribution bar chart
• Fixed version display inconsistency in sidebar
• Rebuilt complete api.ts after a partial-paste regression

v0.9.1 - Historical Price Explorer

• Full price history endpoint with auto-ingest (period=max support)
• Candlestick chart with volume via TradingView's lightweight-charts
• Ticker search on Dashboard works for any valid symbol, not just pre-ingested ones
• 1M/6M/1Y/5Y/All range selector
• Lowered coverage threshold to 65% to reflect growing I/O-heavy real-time surface area

v0.9.0 - Real-Time Price Streaming

• WebSocket endpoint (/ws/prices) with subscriber-based refresh throttling
• Auto-ingest on first request — no more manual POST before viewing a ticker
• Graceful fallback to last known close when markets are closed
• Frontend usePriceStream hook with auto-reconnect
• Overview page ticker tape now shows genuinely live prices
• 130+ tests passing

v0.8.1 - Production Deployment + ML Pipeline Fix

• Deployed to Railway (FastAPI + TimescaleDB + pgvector) and Vercel (Next.js frontend)
• Fixed 52-week high/low feature engineering bug — previously required exactly 252+ rows, now gracefully degrades with min_periods=1, verified at the 251-row boundary in production
• Full agent pipeline (LangGraph + Groq LLM) confirmed working end-to-end live
• CORS wired between Vercel frontend and Railway backend

v0.8.0 - Advanced Algorithms

• Fama-French 3-factor model (alpha, beta, SMB, HML with t-statistics)
• TWAP execution scheduling
• VWAP execution scheduling (U-shaped intraday volume profile)
• Implementation Shortfall scheduling (urgency-parameterised)
• Post-trade execution quality analysis
• ~120 unit tests passing, CI green