◉ Product Blueprint — March 2026

StockPilot: An AI-Native Social Investing Platform

A separate product built on StockEdge's infrastructure — combining its analytical depth with social engagement, gamification, and a Foundation Model-first AI layer. Designed to hook India's next 10 million investors.

12 mo
Full build horizon
3 mo
Phase 1 MVP
FM-First
No traditional ML from day 1
0
Duplicate infrastructure

Product Thesis

Why StockPilot, and Why Now

India's retail investor base is exploding, but the market is split: deep analytics on one side, shallow engagement on the other. StockPilot unifies both — powered by StockEdge's engine.

01 — OPPORTUNITY

The window is narrowing

StockGro is rapidly adding analytical depth. The competitive advantage of building engagement on top of StockEdge's existing data moat has a finite lifespan. Moving now means StockEdge's dual SEBI registration (RA + IA) — which takes 12–18 months for any new entrant — becomes a structural head start for AI feature compliance.

02 — APPROACH

FM-first, not ML-first

Instead of spending months training bespoke ML models, StockPilot starts with Foundation Models (Claude / GPT-4o) wrapped with precomputed statistical features and retrieval-augmented generation. This collapses the AI development timeline from months to weeks, lets us iterate on product-market fit before committing to fine-tuning, and keeps costs manageable through intelligent caching and prompt engineering.

03 — ASSET LEVERAGE

StockEdge is a platform, not a constraint

StockPilot consumes StockEdge like a microservice: scan engine via API adapter, market data via Kafka subscription, SEBI licences via legal umbrella, broker integrations via passthrough. The adapter architecture means zero duplicate infrastructure and independent deployability for both codebases.


System Architecture

Adapter-Based Integration with StockEdge

The new app never imports StockEdge code — it only calls typed adapter interfaces. If StockEdge changes an API, only the adapter changes. Both codebases remain independently deployable.

Adapter Contract StockEdge Provides StockPilot Consumes
ScanEngineAdapter POST /scan/execute 500+ scan logic, filters, execution engine NL scan compiler wraps this API — no rebuild needed
DataFeedAdapter Kafka consumer / REST NSE/BSE real-time ticks, 10yr historical EOD Paper trading pricing, chart data, portfolio valuation
BrokerAdapter REST passthrough Zerodha, Angel One, Kotak, Upstox — live & tested Simulated fills for paper trading engine
AuthAdapter OAuth2 / JWT exchange Existing user DB + JWT issuer SSO — users log in with StockEdge credentials
FinancialDataAdapter REST / GraphQL (read-only) 10yr fundamentals, corporate actions, FII/DII data AI portfolio analysis, FM reasoning context
NewsEventAdapter Kafka event subscription NSE filings, earnings, news pipeline AI earnings summaries, alert triggers, sentiment input
UserProfileAdapter REST (read-only) User profile, KYC state, subscription tier Personalisation, feature gating, pro tier sync
Design principle: Adapters are versioned interfaces with a 4-week deprecation policy. The critical first action is a 2-hour SE engineering API workshop in Week 1 to document all contracts. This single session unblocks ~80% of the build.
Net-new (StockPilot builds)
Consumed from StockEdge
Adapter interface boundary

AI Strategy

Foundation Model-First: AI as the Operating System

Not AI as a feature bolted on — AI as the core interaction layer across every screen. Foundation Models handle reasoning; precomputed statistical models handle data prep. No bespoke ML training in Phase 1.

NL Scan Compiler

Users type natural language queries like "show me profitable small-caps with promoter buying" — the FM translates this into StockEdge's scan DSL and executes via the ScanEngineAdapter. No bespoke NLP model needed; the FM handles intent parsing, the scan engine handles execution.

FM → Scan DSL → SE Adapter

Context-Aware Copilot

Every screen gets an AI layer. On a stock detail page, the copilot synthesises news sentiment (precomputed), technical signals (from SE scans), and fundamental data (from SE data lake) into a plain-English intelligence brief. No recommendation — just structured context the user can act on.

FM + RAG + Precomputed Stats

AI Content Engine

Auto-generates daily market briefs, earnings summaries within 5 minutes of filing events, and sector wrap-ups. Triggered by NewsEventAdapter events. FM generates the narrative; precomputed statistical context (price changes, volume, peer performance) is injected via prompt.

Event-Driven FM Generation

Portfolio Health Scorer

Analyses user's paper (or synced real) portfolio using precomputed diversification metrics, sector concentration ratios, and benchmark deviation — then the FM generates a narrative health report. Rule engine computes the numbers; FM explains them in accessible language.

Rule Engine + FM Narrative

SEBI Guardrail Classifier

Every AI output passes through a classification layer before reaching the user. The FM itself is prompted with guardrails, and a lightweight classifier (FM-based, not a separate model) categorises each output as data, insight, or advice — blocking anything in the advice category.

FM Self-Classification + Prompt Guard

Sentiment Aggregator

Instead of training a custom FinBERT model in Phase 1, we use the FM to score news sentiment in batch (nightly precomputation for cost efficiency). Results feed a Market Mood Index and per-stock sentiment metrics. Upgrade to fine-tuned model only when volume justifies the cost.

FM Batch Scoring → Precomputed Cache

The Cost-Optimised AI Pipeline

Every AI feature follows the same pattern: precompute what's computable (statistical models, rule engines, cached aggregations), then use the FM only for reasoning, synthesis, and natural language generation. This minimises FM API calls while maximising output quality.

SE Data Lake
Fundamentals, prices, filings
Precompute Layer
Stats, ratios, signals, sentiment
Redis/Postgres Cache
Ready-to-inject context
FM Reasoning (Claude)
Synthesis + NL generation
SEBI Guardrail
Classify → data/insight/advice
User
Compliant, contextual output
~₹1.5
Per AI query (avg with caching)
60%
Queries served from cache
Phase 3
Fine-tuned on-premise LLM (if scale justifies)

Why FM-First, Not ML-First

Traditional approach: spend 2–3 months collecting labelled data, training NLP/classification models, deploying inference pipelines. FM approach: ship in 2–3 weeks with prompt engineering, iterate based on real user behaviour, then decide what to fine-tune. The FM acts as a rapid prototyping layer for AI features — commit to bespoke models only when you have distribution and usage data to justify the investment.

When to Upgrade from FM to Custom Model

The decision triggers are: query volume exceeds cost threshold (>₹5L/month on FM APIs), latency requirements tighten below 200ms (FM calls take 1–3s), or domain-specific accuracy on repeated tasks falls below 90%. Until then, FM + prompt engineering + caching is the optimal cost-quality tradeoff for a product finding PMF.


Product Blueprint

What StockPilot Builds — Net-New

With StockEdge handling all data infrastructure, StockPilot's codebase is focused entirely on the engagement layer, AI reasoning, and the Flutter client. Every feature below is net-new.

Engagement Layer — Learn by Doing
Paper Trading Engine
Virtual ₹5L portfolio. Real-time simulated P&L using live tick data via DataFeedAdapter. Simulated fills via BrokerAdapter — no broker approval needed.
Go · Redis (position ledger) · WebSocket (live P&L)
Trading Contests & Leaderboards
Weekly leagues with Redis sorted-set leaderboards. Live rank updates via WebSocket. Contest lobby, history, and virtual prize management. Campus league integration via B2B API.
Go · Redis sorted sets · WebSocket · Campus B2B REST API
AI-Guided Onboarding
FM infers user type (beginner / trader / investor) within 3 interactions and adapts the entire UX. Paper trading + guided first scan + AI-explained results create an immediate value loop.
FM classification · Adaptive UI routing
Gamified Learning Streaks
Daily scan challenges, quiz streaks, and portfolio milestone badges. Progression system tied to increasing analytical complexity — users graduate from basic scans to multi-factor strategies.
Go contest engine · Redis · Notification Dispatcher
Social Layer — Community Intelligence
Social Graph & Feed
Follow users and SEBI-verified experts. Infinite-scroll feed with trade ideas, watchlist updates, and scan results shared by the community. Write-time fanout architecture for feed delivery.
Go · Neo4j/DynamoDB · Write-time fanout
Expert Follow & Portfolio View
SEBI-verified expert accounts with real-time portfolio visibility. Users observe expert strategies and paper-trade alongside them. Copy trading (RA-compliant) enabled in Phase 3 under SE's licence umbrella.
Go · AuthAdapter (SEBI verification) · WebSocket
Public Watchlists & Shared Scans
Users publish watchlists and scan configurations. Community can clone, modify, and share — creating a network effect around StockEdge's 500+ scan engine. Each shared scan is a re-engagement hook.
ScanEngineAdapter · Social Graph Service
Social Intelligence Engine (Phase 3)
FM detects trending strategies, rising expert portfolios, and unusual cluster behaviour across the community. Surfaces these as social signals — not advice. Think "trending on StockPilot" for trade ideas.
FM batch analysis · Redis · Feed ranking
AI Layer — FM-Powered Intelligence
NL Scan Builder
Natural language → StockEdge scan DSL compiler. "Show me IT stocks with rising promoter stake and low debt" translates to multi-filter scan execution. FM handles intent; SE scan engine handles data. Free tier feature — the primary user acquisition hook.
FM (Claude) → Scan DSL → ScanEngineAdapter
AI Daily Market Brief
FM-generated morning brief using NewsEventAdapter and DataFeedAdapter data. Covers index performance, top movers, sector summary, FII/DII activity — all data-only, no recommendations. Delivered via push, WhatsApp, and Telegram.
FM + Precomputed Stats → Notification Dispatcher
Earnings Summary Generator
Auto-triggers on NewsEventAdapter filing events. Within 5 minutes of an earnings release, FM generates a structured summary: revenue vs estimates, margin changes, management commentary highlights, peer comparison. Precomputed financials injected via prompt.
Event-driven · FM + FinancialDataAdapter context
AI Copilot (Phase 3)
Context-aware NL Q&A overlay on every screen. Stock detail page → synthesises news + technicals + fundamentals. Portfolio page → explains concentration risk. Scan page → suggests refinements. All responses pass through the SEBI guardrail classifier.
FM + RAG + Guardrail Classifier
Multi-Channel Notifications
WhatsApp Alerts
Price alerts, scan triggers, earnings notifications, and contest updates via WhatsApp Business API. Applied under StockEdge's verified business entity for faster Meta approval.
WhatsApp Business API · Notification Dispatcher (Go)
Telegram Bot
Advanced flows: portfolio digest, alert configuration, scan result delivery, contest rank updates. 1–2 day integration — no approval required. Phase 2 adds interactive flows (configure alerts via chat).
Telegram Bot API · Go handler · Redis session

SEBI Compliance

The Safe Zone Is Large — and StockEdge Expands It

StockEdge's dual RA (INH300007493) + IA (INA000017781) registration gives StockPilot the widest possible regulatory latitude. Every AI feature is designed around the data-vs-advice boundary.

✓ Safe zone — What StockPilot can build freely

AI-powered stock screeners (NL → filter logic, any metric)
News summarisation + sentiment metrics presented as data
Portfolio analytics: composition, risk, diversification, benchmarks
Technical chart pattern detection (AI-identified, user-interpreted)
Market Mood Index / sentiment heatmaps (data presentation)
AI daily market brief (general, no specific stock tips)
Educational chatbot (explains RSI, P/E, SIP — not trade signals)
Trending stocks / unusual volume / 52-week highs (data screens)
Paper trading (simulated, no real money or advisory element)

✗ Requires RA/IA — Gated under SE licence umbrella

Buy / Sell / Hold recommendations on specific securities
Specific price targets or stop-loss levels
Personalised model portfolios tailored to individual users
Intraday or swing trade calls on stocks
AI-generated copy trade automation
Black-box algo strategies (requires RA + exchange pre-approval)
"AI says you should invest in X" messaging of any kind
Guardrail implementation: The SEBI guardrail classifier is deployed before any AI feature goes live. It's an FM-based classification layer that labels every output as data, insight, or advice. Advice outputs are blocked at the prompt layer. Regulation 24(7) requires disclosure of AI usage to users — every AI-generated output carries a mandatory disclosure tag. SEBI looks at substance over form: disclaimers alone don't provide safe harbour.

Technical Stack

Production Architecture — Built for Indian Scale

Flutter for cross-platform (validated at Zerodha scale: 3M+ daily orders, 2 developers), Go for backend services, Python for the AI/FM layer. Offline-first and low-end device optimised.

Client Layer
Flutter 3.x
Cross-platform, proven at Zerodha scale
BLoC + Riverpod
Stream-native state mgmt + DI
Syncfusion Charts
OHLC, candlestick, built-in indicators
Drift (SQLite)
Offline-first, ACID-compliant local DB
go_router
Deep linking support
WebSocket + throttle
100–500ms tick conflation for UI
Backend Layer
Go (Golang)
Microservices: social, paper trading, contests
Kong API Gateway
Auth middleware, rate limiting, routing
Apache Kafka
SE data/news event consumption
Redis (ElastiCache)
Leaderboards, cache, paper positions
PostgreSQL (RDS)
Primary relational store
Neo4j / DynamoDB
Social graph storage
AI / FM Layer
Claude / GPT-4o
FM gateway for reasoning + NL gen
RAG Pipeline
NSE filings, news, company data retrieval
NL → Scan DSL
FM-based intent parser + compiler
Pinecone / Qdrant
Vector DB for semantic search
Python (FastAPI)
AI service orchestration
Guardrail Classifier
FM-based data/insight/advice filter
Infrastructure
AWS ECS/EKS
Container orchestration (existing SE infra)
GitHub Actions + Codemagic
CI/CD for Flutter + backend
Sentry + Datadog
Error tracking, latency, cost dashboards
Firebase FCM
Push notifications
WhatsApp Business API
Multi-channel alert delivery
DPDP Act compliance
PII encryption, consent management
India-specific optimisations: APK under 30MB (--split-per-abi). ListView.builder with itemExtent for stock lists. RepaintBoundary for live price tickers. Dart Isolates for heavy computation. Data-saver mode reducing tick frequency on 3G/4G. Offline-first with Drift local reads + background sync. Binary protocol parsing (Protocol Buffers) for 60–80% smaller payloads vs JSON.

Dependencies

What StockPilot Needs from StockEdge

All dependencies are internal API agreements — not external contracts or regulatory filings. The risk profile is fundamentally different from a greenfield build.

SE Backend Team

Scan Engine API Docs + Sandbox

Documented REST API spec for /scan/execute. Sandbox access for dev/test. Must be stable and versioned — NL scan builder cannot be tested without this.
HIGH RISK — Blocks NL scan builder
SE Data Team

Data Feed Access

Either Kafka topic subscription or REST endpoint for tick data + historical EOD. Defines paper trading fidelity and latency. Must be settled before Sprint 1.
HIGH RISK — Blocks paper trading
SE Platform Team

Auth / SSO Specification

JWT signing keys or OAuth2 server for SSO. Users log in to StockPilot with StockEdge credentials. Without SSO, users need separate login — hurts activation rate.
HIGH RISK — Blocks onboarding UX
StockEdge Legal

SEBI Licence Umbrella

Written confirmation that StockPilot operates under SE's existing RA/IA licences. Required before any AI features ship to users.
HIGH RISK — Blocks AI feature launch
SE Data Team

Financial Data API

Read-only REST/GraphQL endpoint to 10yr data lake (fundamentals, ratios, FII/DII). Required for AI portfolio analysis and FM context injection.
MEDIUM RISK
SE Data Team

News + Filings Event Stream

Kafka topic or webhook for NSE filings and news events. Needed for AI earnings summary engine and alert trigger pipeline.
MEDIUM RISK
SE Integrations

Broker API Access (Read-Only)

BrokerAdapter must be callable for simulated paper fills. Passthrough OK — no live order routing needed. Paper trading doesn't need broker approval.
MEDIUM RISK
SE Architecture

API Versioning Agreement

Agreement that SE will version internal APIs and give 4-week deprecation notice. Prevents adapter breaks from unannounced SE changes.
MEDIUM RISK
SE Marketing

WhatsApp Business API

Use SE's existing approved WhatsApp Business account, or apply under SE's entity. Faster than a fresh application under a new entity.
LOW RISK

Development Roadmap

3 Phases, 12 Months — MVP in 3

Phase 1 is lean (3 months, not 5) because all data infrastructure is provided by StockEdge. The adapter workshop in Week 1 is the only true blocker.

WEEK 0
Before coding starts
SE Engineering API Workshop — The Only Blocker
Document all 7 adapter contracts in a 2hr session
Agree API versioning + deprecation policy
Set up dev environment + CI/CD baseline
Apply for WhatsApp Business API under SE entity
Confirm SEBI licence umbrella with SE legal
Provision AWS account (under SE organisation)
PHASE 1
Month 1–3
2 eng core + fractional PM/AI/QA
MVP — Paper trading + NL scan + social basics + WhatsApp alerts
Adapter layer: all 7 typed Go interfaces to SE services
Flutter app (iOS + Android): feed, paper trading, scan, alerts
Paper trading engine: ₹5L virtual, Redis positions, sim fills
NL scan builder: text → DSL compiler via ScanEngineAdapter
Social basics: follow, public watchlists, activity feed
WhatsApp + Telegram alerts: price + scan triggers
SEBI guardrail classifier: deployed before AI goes live
AI daily market brief: FM morning brief via adapters
PHASE 2
Month 4–6
+2 engineers (backend/infra)
Contests + expert follow + campus B2B + AI portfolio health
Trading contest engine: weekly leagues, leaderboards, prizes
Expert follow + live portfolio view (SEBI-verified only)
Campus B2B API: REST + white-label for institutions
AI portfolio health score: diversification, risk, benchmark
Earnings AI summary engine: auto-publish within 5min
Sentiment aggregator: FM batch → Market Mood Index
PHASE 3
Month 7–12
Full team scaled
AI copilot everywhere + copy trading + LLM cost optimisation
AI copilot on every screen: context-aware NL Q&A
Personalised AI feed: behaviour-based scan curation
Copy trading (RA-compliant): under SE licence umbrella
Real cash contest prizes: RBI-compliant payment infra
Fine-tuned on-premise LLM: 80% API cost reduction
Social intelligence engine + predictive churn model

Investment

What It Takes to Build Phase 1

AI coding tools (Cursor, Copilot, Claude Code) multiply a small team to ~2.5x output velocity. Phase 1 is deliberately lean — validate PMF before scaling the team.

Role Count Phase Monthly Cost
Flutter Engineer 1 Phase 1+ ₹3–4L
Backend / Go Engineer 1 → 2 Phase 1 → 2 ₹3–4L
Product Manager 0.25 (fractional) Phase 1+ ~₹0.5L
AI / FM Engineer 0.25 (fractional) Phase 1+ ~₹0.5L
DevOps / Infra 0.5 (shared) Phase 1+ Shared with SE
FM API Costs (Claude/OpenAI) Phase 1+ ~₹0.5–1L
AWS Infrastructure Phase 1+ ~₹0.5L
PHASE 1 TOTAL (3 MONTHS)
~₹27L
Engineering + infra + FM APIs
AI coding multiplier: 2 engineers with Cursor + Claude ship at ~5-person velocity

Structural Advantage

What StockGro Cannot Clone

Features can be copied. Infrastructure moats cannot. StockPilot inherits all of StockEdge's structural advantages on day one.

REGULATORY MOAT

Dual SEBI RA + IA Registration

Takes 12–18 months for any new entrant. Enables the widest range of AI-powered features and compliant advisory products. StockPilot operates under this from day 1.

DATA MOAT

10 Years of Proprietary Financial Data

Unmatched data lake — 10yr financials, 250+ marquee investor portfolios, bulk/block deal history, FII/DII flows. AI is only as good as its context window, and this context is irreplaceable.

PRODUCT MOAT

500+ Battle-Tested Scans

The scan engine is a compounding asset. Every new social feature (contests, copy trading, shared watchlists) becomes dramatically more powerful when built on proven scan infrastructure.

DISTRIBUTION MOAT

Elearnmarkets — 450K Learners

A direct acquisition funnel into a social investing platform. Cross-login SSO means these users are day-1 StockPilot users. No competitor has a captive, high-intent education audience of this scale.


Next Steps

The Only Blocker Is a 2-Hour Meeting

Everything else can run in parallel from Day 1. Time to first working prototype: ~3 weeks after the adapter workshop.

Week 1 — Before Coding Starts

01

SE Engineering API Workshop

2-hour session. Document all 7 adapter contracts. Agree versioning policy. This single meeting unblocks ~80% of the build.

02

SEBI Licence Confirmation

SE legal confirms the new app operates under existing RA/IA licences. Written confirmation required before AI features ship.

03

Infra + CI/CD Baseline

Provision AWS account under SE org. Set up GitHub Actions + Codemagic. Begin Flutter app shell + first 2 adapters in parallel.

6 weeks to working prototype. 3 months to MVP. 12 months to full product. StockEdge has everything StockGro was built without. The question isn't whether to build this — it's how fast.