ProjectsAugust 23, 2026

QuantForge — Multi-Agent Market Intelligence & Algorithmic Trading Platform

Overview

QuantForge is a multi-agent market intelligence and algorithmic trading platform for Indian equities and options (NSE/BSE), built on top of Zerodha's Kite Connect API. Rather than a single monolithic strategy engine, the system is composed of 13 independent agents — regime, technical, quant, options, risk, sector, volatility, portfolio, stock, advisor, strategy-research, mutual-fund, and swing — each producing an opinion, with no single agent holding trading authority on its own. The project is currently an actively-evolving MVP (Phase 1, per its own roadmap): paper-trading only, no live execution yet, with Phase 2/3 (fundamental, macro, and sentiment agents, live order routing) explicitly deferred.

Architecture

Agent-Debate, Risk-Veto Design

No single agent can act unilaterally. Every proposed trade passes through a Risk Agent with final veto authority — the same instinct that drives test gating in CI/CD, applied to capital instead of code: nothing ships (or trades) without an independent check that can say no.

The Quality Gate

Every strategy is required to clear four stages, in order, before it can touch real capital:

Backtest → Out-of-Sample Validation → Paper Trading → Risk Review

This mirrors a staged test-promotion pipeline (unit → integration → staging → prod) more than a typical trading bot's "backtest and go live" pattern — a deliberate choice to keep unproven strategies contained.

Data & Execution Layer

Zerodha Kite Connect as the primary broker/data source for live NSE/BSE quotes and the options chain (no free options-chain data source exists, so options analysis requires an active Kite session)
yfinance as an equities-only fallback for pre-login / no-session states
Real-time index quotes migrated from HTTP polling to Kite's WebSocket ticker for lower latency
Postgres + TimescaleDB for time-series storage, Redis for caching (Kafka planned for a later phase)
Explicit, documented inter-agent JSON contracts (docs/SCHEMAS.md) so agents can be developed, tested, and replaced independently

Options Intelligence

A dedicated agent computes implied volatility and Greeks via Black-Scholes and scores strategies by max profit, max loss, and breakeven — surfaced through the dashboard's Options Intelligence view.

Dashboard

A Next.js 16 / React 19 / TypeScript / Tailwind frontend with 8 distinct views: Mission Control, AI Agents, Stock Detail, Opportunities, Options Intelligence, Risk Center, Strategy Lab, and Trade Journal — plus a dual Paper/Real × Manual/Auto trading-mode toggle, with the real-account path built defensively as read-only scaffolding ahead of any live-trading work.

Testing & Quality

64 tests across 10 pytest files covering Black-Scholes pricing, the risk-veto path, backtest look-ahead-bias checks, custom strategy logic, trading-mode switching, and each individual agent (options, mutual fund, stock advisor, strategy research, swing)
Backtest bias checks specifically guard against look-ahead bias — a correctness class of bug that's easy to introduce and easy to miss in trading systems, and directly analogous to flaky-test root causes in E2E suites

Technologies Used

Python 3.12, FastAPI — orchestrator backend (services/orchestrator)
Next.js 16, React 19, TypeScript, Tailwind CSS 4 — dashboard
Postgres, TimescaleDB, Redis — data layer
Zerodha Kite Connect, yfinance — market data and broker integration
pytest — test suite (64 tests, 10 files)
lightweight-charts — dashboard charting

Challenges and Learnings

No single source of truth for options data: Free options-chain data doesn't exist for NSE/BSE, which forced a design where options analysis is only available with an active, authenticated Kite session — equities can degrade to yfinance, options cannot. Keeping agents independently testable: With 13 agents, the JSON-contract-first approach (documented in docs/SCHEMAS.md) turned out to be the difference between agents that can be unit-tested in isolation and a system where every change requires an end-to-end run to verify. Resisting the urge to ship live execution early: The Backtest → OOS → Paper → Risk Review gate is intentionally strict, and the real-account trading path is still read-only scaffolding — a conscious decision to keep the system honest about what it has actually proven versus what it merely computes.

Outcome

QuantForge is currently an MVP (Phase 1): the multi-agent architecture, risk-veto gate, and options intelligence are implemented and covered by 64 tests, running in paper-trading mode. Live execution and additional agent categories (fundamental, macro, sentiment) are scoped for later phases and haven't shipped yet — this case study reflects where the project actually stands, not where it's headed.

More work

Related projects

FlakeIQ — Flake Tracking and LLM-Powered Failure Classification

Built a flake analysis pipeline for mobile E2E tests that captures test context at runtime, classifies failures via Ollama (llama3.2), and visualizes trends in a real-time Chart.js dashboard.

Read case study View project

Cross-Platform Mobile E2E Testing with mobilewright

Configured and maintained a 19-test mobile E2E suite using mobilewright covering alerts, animation, calendar, forms, gestures, lists, media, signature, profile, and login flows — all passing on both iOS and Android. Contributed two upstream bug fixes to the mobilewright framework.

Read case study View project

Adya — From SwiftUI Prototype to Cross-Platform Rewrite

A minimal daily task manager, first validated as a native SwiftUI/iOS prototype, then rebuilt from scratch in Expo/React Native (New Architecture) to ship one codebase across iOS and Android.

Read case study