NIFTY50 ▲ 0.42% 24,318 SENSEX ▲ 0.38% 80,124 RELIANCE ▼ 0.21% 2,941 TCS ▲ 1.14% 3,872 INFY ▲ 0.67% 1,628 HDFCBANK ▼ 0.09% 1,734 WIPRO ▲ 0.55% 478 BAJFINANCE ▲ 1.32% 7,218 ICICIBANK ▲ 0.81% 1,291 NIFTY50 ▲ 0.42% 24,318 SENSEX ▲ 0.38% 80,124 RELIANCE ▼ 0.21% 2,941 TCS ▲ 1.14% 3,872 INFY ▲ 0.67% 1,628 HDFCBANK ▼ 0.09% 1,734 WIPRO ▲ 0.55% 478 BAJFINANCE ▲ 1.32% 7,218 ICICIBANK ▲ 0.81% 1,291
Personal Research Project

FinAgent Multi-Agent Investment Intelligence System

A personal research project exploring how multi-agent AI architectures can be applied to Indian equity markets — combining real-time market data, fundamental analysis, technical signals, and risk modelling into a coordinated intelligence layer.

Built with Claude Code as the agentic backbone, Google Colab as the compute environment, and XTS APIs for live market connectivity. This is an exploration of applied AI engineering — not financial advice.

Claude Code Google Colab XTS Market APIs Python Multi-Agent NSE / BSE
Work in Progress · Private Repository
PORTFOLIO SIGNAL · NIFTY50
↑ 3.2%
▲ Agents consensus: BUY
Simulated output
Not live trading
FUNDAMENTAL AGENT
BUY 78
TECHNICAL AGENT
BUY 65
SENTIMENT AGENT
HOLD 52
RISK AGENT
RISK 28
System Design

Data flow architecture.

A four-layer pipeline: live market data ingestion → specialised AI agents → orchestration → decision output. Each agent operates independently and reports to a central orchestrator.

── LAYER 1 · DATA SOURCES ──
📡
XTS APIs
Live quotes · Order book
→
📰
News & Filings
NSE · BSE · RSS feeds
→
📊
Historical Data
OHLCV · Fundamentals
→
🌐
Macro Signals
RBI · FII/DII flows
↓   ↓   ↓   ↓
── LAYER 2 · GOOGLE COLAB · PYTHON RUNTIME ──
🔄
Data Normaliser
Pandas · Cleaning
→
📐
Feature Engineering
TA-Lib · Indicators
→
🗄️
Context Store
SQLite · Vector DB
↓   ↓   ↓
── LAYER 3 · CLAUDE CODE · MULTI-AGENT LAYER ──
🔍
Fundamental Agent
P/E · EPS · Debt analysis
→
📈
Technical Agent
RSI · MACD · Patterns
→
🧠
Sentiment Agent
NLP · News scoring
→
⚠️
Risk Agent
VaR · Drawdown · Beta
↓   ↓   ↓   ↓
── LAYER 4 · ORCHESTRATOR + OUTPUT ──
🎯
Orchestrator Agent
Claude · Signal synthesis
→
📋
Research Report
Auto-generated PDF
→
🔔
Alerts & Signals
Telegram · Email
→
📊
Dashboard
Colab · Streamlit
Agent Design

Five specialised AI agents.

Each agent has a narrow focus and a well-defined context window. They communicate through structured outputs to a central orchestrator that synthesises a final recommendation.

01
🔍
Fundamental Agent
Value & Quality Analysis
Evaluates P/E ratios, EPS growth, debt-to-equity, promoter holding, and quarterly result trends. Scores stocks on fundamental quality against sector peers.
NSE FilingsScreenerPandas
02
📈
Technical Agent
Price & Momentum Analysis
Analyses RSI, MACD, Bollinger Bands, moving averages, volume patterns, and support/resistance levels. Detects breakout and reversal patterns.
XTS APIsTA-LibOHLCV
03
🧠
Sentiment Agent
News & Social Intelligence
Processes financial news, exchange announcements, and analyst reports using NLP. Generates a sentiment score that flags positive/negative catalysts.
Claude NLPRSSBSE Filings
04
⚠️
Risk Agent
Portfolio Risk Management
Computes Value-at-Risk, beta, max drawdown, sector concentration, and correlation matrix. Flags position sizing constraints before signals are acted on.
SciPyNumPyVaR models
05
🎯
Orchestrator Agent
Signal Synthesis & Decision
The master Claude agent. Receives structured outputs from all four agents, resolves conflicts, applies weighting logic, and produces a final recommendation with reasoning trace.
Claude CodeTool useChain-of-thought
Technical Stack

Built with the right tools.

Technology choices driven by accessibility, cost, and production-readiness — designed to be portable from a Colab notebook to a cloud deployment.

AI & Orchestration
🤖
Claude Code (Anthropic)
Agentic backbone — tool use, multi-step reasoning, structured outputs, chain-of-thought orchestration
🔗
Multi-Agent Framework
Custom orchestrator pattern with specialised sub-agents and context handoffs
🧮
LLM-as-Judge
Using Claude to evaluate and score agent outputs before synthesis
Data & Market Connectivity
📡
XTS Interactive APIs
Real-time quotes, order book depth, historical OHLCV, market depth for NSE/BSE
📊
TA-Lib + Pandas
Technical indicator computation, data normalisation and feature engineering
🗄️
SQLite + Vector Store
Persistent context store for agent memory and semantic similarity search
Compute & Infrastructure
☁️
Google Colab
Primary compute environment — GPU access, scheduled notebooks, Drive integration
🐍
Python 3.11+
Core language — asyncio for concurrent agent execution, type-annotated throughout
📬
Telegram Bot + Email
Signal delivery layer — real-time alerts with full reasoning context
Quality & Reliability
✅
Golden Dataset Testing
Historical signal backtesting against known market events to validate agent accuracy
🔄
A/B Signal Testing
Parallel agent configurations run against the same data to compare recommendation quality
📝
Reasoning Traces
Every agent decision is logged with full chain-of-thought for auditability
Engineering Approach

Principles that guide the build.

24 years of quality engineering applied to AI system design — the same discipline, a new domain.

01
Narrow agents, wide system
Each agent has a single responsibility. Complexity emerges from composition, not from individual agents trying to do everything.
02
Reasoning over rules
Agents reason about context rather than follow hardcoded rules. The system adapts to market regime changes without manual tuning.
03
Auditability first
Every signal carries its full reasoning trace. Drawing from QA roots — if you can't explain a decision, you can't trust it in production.
04
Risk before reward
The Risk Agent is a hard gate. No signal reaches the orchestrator without passing risk constraints — position sizing beats signal quality.
05
Golden dataset validation
Agent outputs are continuously validated against historical events where correct signals are known — the same pattern as QA efficacy labs.
06
Research, not automation
This is a decision-support system. Final trading decisions remain human. The system surfaces insights; the human acts on them.
Currently in active development
Private repository · Personal research project · Not financial advice
5
AI Agents
4
Data Sources
NSE
Market
WIP
Status