Sensei AI — Institutional-Grade Stock Analysis Platform
Budget / SalaryHourly project
TypeFreelance project
LocationRemote
Posted1 hour ago
Sensei AI — Institutional-Grade Stock Analysis Platform
I built Sensei AI, a real-time trading intelligence platform covering all 50 Nifty stocks, designed to bring institutional-level AI analysis to retail investors.
The system unifies five distinct AI model families into a single scored decision engine (−5.0 to +5.0):
- Classical ML — Random Forest with SHAP-based feature attribution for explainable UP/DOWN probability
- Deep Learning — Custom LSTM and causal Temporal CNN (PyTorch) for 5-day return forecasting
- Reinforcement Learning — PPO agent (Stable-Baselines3) trained via a custom Gymnasium trading environment for BUY/SELL/HOLD actions
- Regime Detection — Gaussian Hidden Markov Model to classify BULL/BEAR market states
- Financial NLP — FinBERT sentiment analysis on live Google News RSS feeds
These signals feed a weighted voting decision engine that outputs a confidence-scored BUY/SELL/HOLD call, backed by a 5-method Support & Resistance engine, automated
intraday/swing trade setup generation (Entry/SL/Target/R:R), and backtesting (Sharpe Ratio, Max Drawdown, Total Return).
The platform ships as a full-stack Streamlit web app with a FastAPI backend, containerized with Docker, and pulls live market data via yfinance and nsepython.
Tech stack: Python 3.13, PyTorch, Stable-Baselines3, Scikit-learn, SHAP, HuggingFace Transformers, hmmlearn, Streamlit, Docker
I built Sensei AI, a real-time trading intelligence platform covering all 50 Nifty stocks, designed to bring institutional-level AI analysis to retail investors.
The system unifies five distinct AI model families into a single scored decision engine (−5.0 to +5.0):
- Classical ML — Random Forest with SHAP-based feature attribution for explainable UP/DOWN probability
- Deep Learning — Custom LSTM and causal Temporal CNN (PyTorch) for 5-day return forecasting
- Reinforcement Learning — PPO agent (Stable-Baselines3) trained via a custom Gymnasium trading environment for BUY/SELL/HOLD actions
- Regime Detection — Gaussian Hidden Markov Model to classify BULL/BEAR market states
- Financial NLP — FinBERT sentiment analysis on live Google News RSS feeds
These signals feed a weighted voting decision engine that outputs a confidence-scored BUY/SELL/HOLD call, backed by a 5-method Support & Resistance engine, automated
intraday/swing trade setup generation (Entry/SL/Target/R:R), and backtesting (Sharpe Ratio, Max Drawdown, Total Return).
The platform ships as a full-stack Streamlit web app with a FastAPI backend, containerized with Docker, and pulls live market data via yfinance and nsepython.
Tech stack: Python 3.13, PyTorch, Stable-Baselines3, Scikit-learn, SHAP, HuggingFace Transformers, hmmlearn, Streamlit, Docker
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.