NeuralQuant

AI-Powered Stock Intelligence Through Adversarial Multi-Agent Debate

Institutional-grade quant engine + PARA-DEBATE multi-agent system + live voice analysts. US + India. 100% live data.

This codebase is available for acquisition. Contact the author for details.

Live at neuralquant.co · API neuralquant.onrender.com (v4.1.3)

---

Current Status — September 2026

Production live. API v4.1.3, ~950 stocks in the score cache (≈450 US + ≈500 India), nightly refresh healthy. Last full prod smoke (2026-08-18): 15/15 endpoints PASS, zero JS console errors across the app.

Recently shipped (Sessions 90–109):

• Branding unified to NeuralQuant; landing, methodology, pricing refreshed.
• Voice analysts live — QuantAstra (on-demand) + Veronica (page-aware companion with "Hey Veronica" wake word + morning briefing), via a LiveKit worker. TTS migrated from ElevenLabs to LiveKit Inference (cartesia/sonic-3.5).
• Hermes — autonomous self-improving trading agent surfaced at /hermes (live "Matrix" dashboard, SSE log stream). Migrated from expired Railway trial to a GCP Always Free e2-micro VM.
• India data parity — QuantFactor universe expanded from ~60 to ~500 NSE names via a daily-refreshed sheet sync. Legacy anjali_enrichment collapsed onto quantfactor_universe.
• Security hardening (P0–P6) — Supabase RLS, log redaction, gitleaks + dependency scanning in CI, IDOR fixes, HTTP security headers + CSP (report-only), per-IP abuse limiting, webhook-signature enforcement, a security-event audit log, and an incident-response runbook. See Security.
• Score-scale bug family fixed — composite_score normalization that broke trade Kelly sizing, meta scores, and backtest hit rate.
• GCP India price feeder — Render's outbound IPs are blocked by Yahoo, so a self-contained feeder on the GCP VM now fetches Indian prices and writes them to stock_snapshot.
• Screener — market=BOTH/GLOBAL merges US + IN; empty/500 bugs fixed.
• QA pass — fixed false "loss-making" badge, CSP console error, "Sign In while authenticated" nav bug, null change_pct, broken SSE stream, empty stock_snapshot cold-start; removed the unused /alerts page and all email functionality.
• Mobile — hamburger nav + bottom tab bar + responsive cards/charts/tables across the app; tightened cramped score grids on phones (verified live at 390px).

Operator follow-ups (not code): manual-deploy quantastra-agent Render worker if not yet deployed, final deletion of any remaining Railway account. The GCP India price feeder is now functional. Before any live demo, run python scripts/warmup.py (~3–5 min prior) — first-hit Ask Morgan / PARA-DEBATE is ~50s / ~85s of inherent multi-agent LLM time, and warming the exact demo tickers makes the audience-facing clicks fast.

---

Acquisition Overview

AI-powered stock intelligence platform — India + US.

Multi-agent Claude-powered research engine (PARA-DEBATE) · QuantFactor quintile engine + proprietary IRS% score across ~950 stocks · web + mobile (Expo) + voice (LiveKit) + an autonomous trading agent on one FastAPI backend · Anthropic Claude with an AWS Bedrock path. See neuralquant.co/methodology for the research approach.

18+ months of build. 90+ documented sessions.

Evaluate in 15 minutes — zero API keys

git clone <repo> && cd NeuralQuant
cp .env.example .env          # DEMO_MODE=true is the docker-compose default
docker compose up --build
# → http://localhost:3000

The demo stack bundles a real data snapshot (Supabase tables, ~1,669 rows) and a canned PARA-DEBATE output, so the dashboard, screener, stock detail, and backtest pages work offline. Add ANTHROPIC_API_KEY to run live agent debates.

What's included in the sale

Codebase · neuralquant.co domain · Supabase data + Q1FY27 backtest results · scoring methodology IP (IRS%, G Score, Risk Efficiency, QuantFactor quintiles) · sister repo (anjali-value-stocks) · brand/design system · optional 3–6 month founder transition. Full list: docs/ASSET_INVENTORY.md. Due diligence: docs/ARCHITECTURE.md · docs/BUG_HISTORY.md · docs/OPERATIONS.md · docs/HANDOVER.md. Patent strategy: IP_PROTECTION/MEETING_PREP.md.

---

What is NeuralQuant?

A full-stack AI stock intelligence platform that combines a quantitative signal engine with a multi-agent AI analyst debate (PARA-DEBATE), natural-language and voice analysts, and an autonomous trading agent — backed entirely by live, real data.

Why It's Different

Feature	NeuralQuant	Danelfin	TipRanks	Seeking Alpha
Adversarial agent	Mandatory BEAR challenge	No	No	No
India market	~500 NSE names, live	No	No	No
Natural language queries	Yes (Ask Morgan)	No	No	No
Voice analyst	Yes (QuantAstra + Veronica)	No	No	No
LLM hallucination guard	Per-agent metric reconciliation vs live data	No	No	No
Real-time enrichment	RSI/MACD/ATR live	End-of-day	End-of-day	End-of-day
Multi-agent debate	6 agents + head analyst, 1 verdict	Single score	Aggregation	Single author
Pricing	Free / $9.99 / $29.99 / $99.99	Free / $20 / $49 / $99	Free / $30 / $50	$239/yr

---

Key Features

• PARA-DEBATE — 5 specialist agents (Macro, Fundamental, Technical, Sentiment, Geopolitical) + adversarial BEAR challenge + Head Analyst synthesis, run in parallel (asyncio.gather). Each agent's numeric claims are reconciled against live authoritative data before synthesis (hallucination guard). SSE-streamed.
• Ask Morgan — natural-language queries with conversation memory and live data injection. "Is AAPL a buy?" → price, P/E, IRS%, QuantFactor breakdown, consensus.
• Voice analysts — QuantAstra (on-demand voice) and Veronica (page-aware companion: wake word, morning briefing, reads the page you're on). Deepgram → Claude → LiveKit Inference pipeline on a LiveKit worker.
• QuantFactor engine + IRS% — cross-sectional quintile scoring (Growth / Return / Valuation / Risk) vs index peers, distilled into the proprietary Investment Readiness Score (IRS%). US + India.
• Hermes — autonomous self-improving trading agent (paper). LLM "reflection" loop mutates one strategy variable at a time with a written rationale and version history. Surfaced at /hermes.
• Screener — filter/rank by IRS%, QuantFactor composite, sector, market; presets.
• Backtest — walk-forward research tooling. No backtest or performance figures are published.
• Dual market — US (S&P 500) + India (NIFTY 500 pool) with India-specific signals (delivery %, India VIX regime).
• Payments — Stripe + PayPal subscriptions, 4 tiers. Free during the development phase.

---

Architecture

┌──────────────────────────────────────────────────────────────────────┐
│                  Next.js 16 Frontend (Obsidian Quantum)               │
│  Dashboard · Screener · Stock Detail · Ask Morgan · Portfolio ·       │
│  Watchlist · Hermes · Backtest · Methodology · Veronica (global orb)  │
└───────────────┬───────────────────────────────────────┬──────────────┘
                │ REST / SSE (via /api proxy)            │ LiveKit (wss)
┌───────────────▼───────────────────────────────────┐   │
│                  FastAPI Backend (nq-api)          │   │
│  34 routers · multi-agent debate · 4-tier cache    │   │
│  Security: JWKS auth · RLS · ADMIN_EMAILS gate ·   │   │
│            IP abuse limiter · webhook sig verify ·  │   │
│            log redaction · security_events audit    │   │
│                                                     │   │
│  Stocks/Screener/Analyst/Query/Market/Portfolio ·   │   │
│  Auth/Checkout(Stripe+PayPal)/Webhooks ·            │   │
│  Watchlists/Backtest/Sentiment/Hermes(proxy)/Team   │   │
└──────┬───────────────────┬────────────────┬────────┘   │
       │                   │                │            │
┌──────▼─────┐  ┌──────────▼────────┐  ┌────▼─────────┐  │
│ nq-signals │  │ PARA-DEBATE        │  │ QuantFactor  │  │
│ 5-factor + │  │ 6 agents ∥ +       │  │ quintile eng │  │
│ HMM regime │  │ head analyst       │  │ + IRS% (US/IN│  │
└──────┬─────┘  │ + metric reconcile │  │  daily sync) │  │
       │        └────────────────────┘  └──────────────┘  │
┌──────▼──────────────────────────────┐   ┌───────────────▼──────────────┐
│            nq-data layer             │   │   livekit-agent (worker)     │
│ yfinance · FMP · NSE Bhavcopy · FRED │   │  QuantAstra + Veronica       │
│ EDGAR F4 · Finnhub · OpenBB proxy    │   │  Deepgram→Claude→Inference   │
└──────────────────────────────────────┘   └──────────────────────────────┘

External: Hermes trading agent (GCP VM) ──proxied via /hermes──▶ nq-api

Apps: apps/web (Next.js, Vercel) · apps/api (FastAPI, Render) · apps/livekit-agent (voice worker, Render) · apps/team (internal Team Hub, Vercel) · apps/mobile (Expo). Packages: packages/data (nq-data), packages/signals (nq-signals).

---

Live Endpoints (Verified 2026-08-18 — 15/15 PASS)

Endpoint	Method	Status	Notes
/health	GET	200	v4.1.3, ~950 cache rows
/stocks/AAPL?market=US	GET	200	IRS%, ForeCast, factor radar
/stocks/TCS?market=IN	GET	200	India full render incl ₹ prices
/stocks/AAPL/chart?period=1mo	GET	200	OHLCV
/stocks/AAPL/meta	GET	200	Market cap, P/E, beta
/screener/preview?market=US	GET	200	Ranked results
/market/overview	GET	200	4 indices
/market/sectors	GET	200	11 sectors
/market/news	GET	200	Headlines + sentiment
/market-wrap/today	GET	200	Daily wrap
/query/v2 (Ask Morgan)	POST	200	~45s cold, fast warm
/analyst/stream (PARA-DEBATE)	POST	200	SSE, ~77s cold
/hermes/status	GET	200	Trading agent state
/auth/me	GET	401	Bearer required

---

Data Sources

Every number is sourced from live APIs — no synthetic data.

Data Point	Source	Refresh
VIX / India VIX	yfinance ^VIX / ^INDIAVIX	Real-time
S&P 500 vs 200-day MA	yfinance ^GSPC	Real-time
HY Credit Spread OAS	FRED BAMLH0A0HYM2	Daily
Profile / quote / fundamentals	FMP (primary) → yfinance (fallback)	Real-time
QuantFactor quintiles (IN)	Daily "NSE 100/500 Analysis" sheet sync	Daily
P/E TTM / P/B ratio	FMP / yfinance	Real-time
Delivery % (India)	NSE Bhavcopy	Daily
Insider buying/selling	SEC EDGAR Form 4	As filed
RSI-14, MACD, ATR-14, SMA	Finnhub candles	15 min
News sentiment	FMP → Finnhub → yfinance cascade	30 min
Extended / terminal data	OpenBB proxy (nq-openbb)	On demand

---

The 5-Factor Model (nq-signals)

Scores are computed cross-sectionally within a reference universe — no hardcoded thresholds.

US Market:

Factor	Signal	Weight (Risk-On)
Quality	Gross margin + Piotroski F-score + Accruals	25%
Momentum	12-1 month price return (Jegadeesh-Titman)	25%
Value	Inverse of (P/E rank x 0.5 + P/B rank x 0.5)	10%
Low Volatility	Inverse of realized 1Y vol + beta rank	15%
Short Interest	Inverse of short float rank	10%
Insider	EDGAR Form 4 cluster score	5%

India Market swaps Short Interest for Delivery % (institutional conviction via delivery_pct rank). Regime weights shift automatically — US uses a 4-state HMM; India uses India VIX heuristics (VIX > 25 = Bear, > 18 = Late-Cycle, else Risk-On).

Alongside this, the QuantFactor engine scores each name into quintiles (Growth / Return / Valuation / Risk) vs index peers and derives the IRS% (Investment Readiness Score) shown across the product.

---

Security

Hardened across a P0–P6 program (Sessions 92–93):

• Auth — Supabase JWT verified via JWKS; admin surfaces gated on an ADMIN_EMAILS allowlist (not tier); internal Team Hub behind admin/service-token.
• Database — Row-Level Security (migrations 026_enable_rls.sql, 027_security_events.sql); backend uses service_role with RLS as the net.
• Secrets — log redaction filter (scrubs keys/tokens/emails); gitleaks secret scanning in CI; .env gitignored.
• Web — HTTP security headers (HSTS, X-Frame-Options, nosniff, Referrer-Policy, Permissions-Policy) + Content-Security-Policy (report-only, pending enforce).
• Abuse / integrity — per-IP rate fuse on expensive endpoints (e.g. LiveKit token); Stripe/PayPal webhook signatures verified (fail-closed); file-upload size caps + MIME allow-list; prompt-injection guards on LLM file analysis.
• Supply chain — pip-audit + npm audit + Dependabot in CI.
• Observability — security_events audit log + incident-response runbook (docs/SECURITY_INCIDENT_RESPONSE.md).

Full audit: docs/SECURITY_IDOR_AUDIT.md · operator actions: docs/SECURITY_P0_P1_OPERATOR_ACTIONS.md.

---

Pricing

Tier	USD/mo	INR/mo (approx)	Features
Free	$0	₹0	PARA-DEBATE, screener, market data
Investor	$9.99	~₹899	Unlimited PARA-DEBATE, Ask Morgan
Pro	$29.99	~₹2,499	+ Watchlists, backtest, portfolio
API	$99.99	~₹8,499	Full API access

All charges in USD via Stripe/PayPal. Quota enforcement is bypassed during the development phase (free indefinitely until monetization).

---

Repository Layout

stockpredictor/
├── packages/
│   ├── data/                  # nq-data: connectors, DataBroker, DuckDB store
│   └── signals/               # nq-signals: 5-factor engine, HMM regime, ranker
├── apps/
│   ├── api/                   # FastAPI backend (34 routers)
│   │   ├── src/nq_api/
│   │   │   ├── auth/              # JWT, admin gate, rate_limit, abuse_limit, security_audit
│   │   │   ├── agents/            # PARA-DEBATE agent prompts + context builders
│   │   │   ├── cache/             # score cache, quantfactor cache, enrichment cache
│   │   │   ├── jobs/              # quantfactor_sync, nightly_score, market_refresh
│   │   │   ├── services/          # stock_summary, portfolio, enrichment, clarification
│   │   │   └── routes/            # stocks, screener, analyst, query, hermes, team, ...
│   │   └── migrations/           # SQL migrations (… 020 RLS, 021 security_events)
│   ├── web/                   # Next.js 16 frontend (Vercel)
│   ├── livekit-agent/         # QuantAstra + Veronica voice worker (Render)
│   ├── team/                  # Internal Team Hub (Vercel)
│   └── mobile/                # Expo mobile app
├── scripts/                   # smoke_test.py, exports, backtest, db backup
├── .github/workflows/         # ci, secret-scan (gitleaks), dep-audit, market-refresh, ...
└── render.yaml                # Render service definitions (api + workers)

---

Quickstart

Prerequisites

Python 3.12+ · uv · Node.js 20+ · Anthropic API key (PARA-DEBATE/Ask Morgan) · FRED + Finnhub + FMP keys (data).

1. Clone and install

git clone https://github.com/satyamdas03/NeuralQuant.git
cd NeuralQuant
uv sync

2. Configure (apps/api/.env)

ANTHROPIC_API_KEY=sk-ant-...
FRED_API_KEY=...
FINNHUB_API_KEY=...
FMP_API_KEY=...
SUPABASE_URL=...           SUPABASE_SERVICE_ROLE_KEY=...   SUPABASE_ANON_KEY=...
[email protected]
# Payments (optional): STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, PAYPAL_*
# Voice (optional, on the livekit-agent worker): LIVEKIT_*, DEEPGRAM_API_KEY (TTS via LiveKit Inference)
# USE_BEDROCK=true to route Claude calls through AWS Bedrock

3. Run

# API
cd apps/api && ../../.venv/bin/uvicorn nq_api.main:app --reload --port 8000
# Web
cd apps/web && npm install && npm run dev    # http://localhost:3000

API at http://localhost:8000 · Swagger at /docs.

4. Tests + smoke

uv run pytest apps/api/tests/ packages/ -v
python scripts/smoke_test.py --api https://neuralquant.onrender.com   # 15 endpoints
python scripts/warmup.py --tickers AAPL,MSFT,TCS                      # pre-demo warm-up

---

Deployment

Service	Platform	URL / Notes
API (nq-api)	Render (Docker)	neuralquant.onrender.com — auto-deploy enabled; verify /health after push
Web	Vercel	neuralquant.co
Voice (quantastra-agent)	Render (worker)	LiveKit; TTS via LiveKit Inference (no separate key)
OpenBB (nq-openbb)	Render	Terminal/extended data proxy
Team Hub	Vercel	Internal ops board
Hermes	GCP e2-micro VM	External trading agent, proxied via /hermes
Database	Supabase	Postgres + Auth + RLS
Payments	Stripe + PayPal	Subscriptions + verified webhooks
CI/CD	GitHub Actions	lint, gitleaks, dep-audit

Scheduling runs via an in-process scheduler (02:00 / 02:30 / 20:30 UTC) — nightly score refresh, QuantFactor sync, EOD market wrap.

---

Competitive Position

Independent expert assessment (May 2026): Overall 8.6/10, USP Distinctiveness 9.0/10.

Dimension	Score	Meaning
Technical Depth	9.2	Academic quant rigour, PARA-DEBATE, zero synthetic data
USP Distinctiveness	9.0	Dual-market + adversarial AI + voice + metric reconciliation
Market Timing	8.4	India 90M+ Demat accounts, AI cost deflation
GTM Readiness	7.8	Product live, acquisition channels needed
Investor Readiness	7.2	Pre-revenue; backtest documented, not multi-year

Market opportunity: $47B TAM (global retail fintech analytics), $3.2B SAM (India + US AI tools), $48M SOM.

---

Roadmap

Completed

• Quant signal engine (5 factors, HMM regime) + QuantFactor quintile engine + IRS%
• PARA-DEBATE (parallel agents + metric reconciliation) + SSE streaming
• 100% live data (FMP, yfinance, FRED, NSE, EDGAR, Finnhub, OpenBB)
• Auth, tiers, screener, portfolio, watchlist, Stripe + PayPal
• Voice analysts (QuantAstra + Veronica)
• Hermes live trading dashboard
• India data parity (~500 NSE names)
• Security hardening P0–P6 (RLS, headers, rate-limiting, audit log, CI scanning)

In progress / next

• Flip CSP from report-only to enforce (after collecting violation reports)
• Cut PARA-DEBATE / Ask Morgan cold-start latency (keep-warm)
• Multi-year backtest with published P&L / Sharpe / IC
• Fitted HMM for India regime (currently heuristic)
• Test coverage expansion
• Populate growth_percentile in score_cache (currently 0.5 default)
• Add /market/indices endpoint for live benchmark quotes

---

Known Limitations

1. Cold-start / inherent LLM latency — first-hit Ask Morgan ~50s, PARA-DEBATE ~85s (multi-agent live debate; faster once warm). Run scripts/warmup.py before live demos.
2. Render auto-deploy — now enabled; verify /health after push for version/cache_age.
3. CSP report-only — security headers shipped; CSP not yet enforced (no violation collector yet).
4. India fundamentals gaps — some FMP endpoints are US-only; IN relies on the sheet sync + yfinance.
5. QuantFactor "earnings declined" flag — derived from profit-growth, not profit level; corrected in UI, full DB correction needs a re-sync.
6. No persistent Ask Morgan memory — sliding-window context, lost on refresh.

---

License

Proprietary. See PROPRIETARY-LICENSE.txt. This codebase is offered for acquisition; not open source.

---

Quant Engine + IRS% · PARA-DEBATE Multi-Agent Debate · Voice Analysts · Autonomous Trading Agent · India & US · 100% Live Data

neuralquant.co · GitHub · API Docs