TradeVision AI: A Multi-Modal Neuro-Stochastic Framework for Real-Time Equity Intelligence Using Brownian Bridge Dynamics and Single-Pass LLM Reasoning
| # | Author | Affiliation | |
|---|---|---|---|
| 1 | Krish Hingu | Department of Computer Engineering, Shri Bhagubhai Mafatlal Polytechnic and College of Engineering, Mumbai | krish.hingu064@svkmmumbai.onmicrosoft.com |
| 2 | Krishna Bhundiya | Department of Computer Engineering, Shri Bhagubhai Mafatlal Polytechnic and College of Engineering, Mumbai | krishna.bhundiya051@svkmmumbai.onmicrosoft.com |
| 3 | Parv Khichadia | Department of Computer Engineering, Shri Bhagubhai Mafatlal Polytechnic and College of Engineering, Mumbai | parv.khichadiya012@svkmmumbai.onmicrosoft.com |
Project Supervisor: Prof. Prachi Arora — Senior Lecturer, Department of Computer Engineering, SBMP & CoE, Mumbai (prachi.arora@sbmp.ac.in)
Every trading day retail traders are dealing with the hodgepodge of data on equity exchanges in India (NSE & BSE). Traditional broker apps separate candlesticks from the streams of real news, so that users are required to make manual connections between technical oscillators and news and corporate developments moving at a lightning pace. The non-linear patterns in prices can be learned by deep neural networks, but the weights of those networks are hard to explain, and require a whole lot of compute that mobile chips aren't capable of.
We introduce TradeVision AI, an end-to-end solution where we combine server-side momentum calculations with the FinBERT sentiment scoring of news with single-pass reasoning using a language model within a mobile Flutter app. To eliminate synthetic chart drift at the discrete market bars, we are creating a stochastic anchoring algorithm for rendering between market bars, directly linking the sub-minute render trajectories to the live exchange quotes (Brownian Bridge stochastic anchoring algorithm, B(t) = W(t) − tW(1) with dual-harmonic wave adjustments).
Our FastAPI back end serializes technical metrics and sentiment polarity into a validated JSON schema. This structured payload is analysed by a hosted LLM during one forward inference pass, with the results including calls to BUY, HOLD and SELL, as well as a calibrated confidence level (50%–95%) and a plain language technical explanation.
Across 252 market sessions on benchmark Indian equities, TradeVision AI attained a 5-day directional accuracy of 67.8% and a Macro F1-score of 0.69, exceeding baseline models by +9.1% while sustaining 59.8 FPS mobile charting.
Stock market forecasting · Explainable AI (XAI) · Large Language Models · FinBERT · Brownian Bridge · Technical Indicators · Flutter · FastAPI · Mobile Computing
| Layer | Technology |
|---|---|
| Frontend | Flutter (Dart) — Cross-platform mobile & web |
| Backend | FastAPI (Python) — REST + WebSocket |
| AI/ML | Groq LLM, FinBERT, XGBoost, Calibrated ML Pipeline |
| Data Sources | Yahoo Finance, Google News RSS, NSE/BSE Live Feeds |
| Database | Supabase (PostgreSQL) |
| Charting | Brownian Bridge Stochastic Anchoring Algorithm |
| Metric | Value |
|---|---|
| 5-Day Directional Accuracy | 67.8% |
| Macro F1-Score | 0.69 |
| Improvement over Baselines | +9.1% |
| Mobile Charting FPS | 59.8 FPS |
| Market Sessions Tested | 252 |
TradeVision/
├── project/
│ ├── backend/ # FastAPI backend (Python)
│ ├── frontend/ # Flutter frontend (Dart)
│ ├── docs/ # Documentation & Research Paper
│ │ └── RESEARCH_PAPER.md ← You are here
│ └── README.md # Project README
© 2026 Krish Hingu, Krishna Bhundiya, Parv Khichadia — All Rights Reserved.