v1.0 · April 2026

Jugalbandi Architecture

Wealth management has a context problem, not a technology problem. The RM has portfolio data. The client has a life. Neither side sees the other. This is the architecture that connects them.

The client dashboard is a signal-collection instrument disguised as a portfolio viewer. The RM command center is a context engine disguised as a task list. They share one Intelligence Graph.

3Signal Types

47AI Functions

4Custodians

7Packages

2Apps

13DB Tables

The Intelligence Graph

A living, per-client model that combines three signal types no other system connects

Every wealth platform has hard data. Positions, NAVs, transactions. That's table stakes. The Intelligence Graph is what happens when you add the two signal types no custodian provides: what the client does on their dashboard, and what the RM knows about the client's life.

Hard Data

Custodians, brokers, PMS, Account Aggregator

Positions across CAMS, KFintech, NSDL, CDSL
NAV, units, avg buy price, XIRR per holding
Transaction history, SIP schedules, dividends
FD details, NPS balance, bank statements via AA

Available everywhere. Not the novel part.

Behavioral Signals

Client dashboard telemetry — silent, continuous

Checked pharma holdings 6x between 2–5 AM
Didn't open app for 11 days after a 4% drawdown
Searched 'SGB maturity' three times in one week
Exported portfolio PDF the day after a market fall

This is the novel signal. No custodian has it.

RM-Contributed Context

RM notes, meeting prep, post-call observations

"Daughter's wedding March 2026 — needs ₹2Cr liquid"
"Doesn't trust structured products since 2008"
"Responds well to data-backed reassurance, not optimism"
"Patriarch makes all decisions. Son is a formality."

Human judgment, encoded. Persists across RM turnover.

These three signals combine into one per-client model

Hard Data

Behavioral Signals

RM Context

Client Intelligence Graph

Without the graph

“Call Mr. Sharma. His portfolio dropped 3.2%.”

With the graph

“Mr. Sharma has checked his healthcare holdings 4 times since Monday. He hasn't called you. Last time this pattern occurred (Sept 2024), he panic-sold Biocon and regretted it in two weeks. His daughter's medical school fees are partially funded from those holdings. He responds well to data-backed reassurance, not general optimism.”

The second briefing reads like a senior colleague's note. It is produced entirely by the system, from signals the RM never had to think about collecting. That is the only design goal of the Intelligence Graph.

Data Pipeline

How custodian data flows into the graph — four integrations, one schema, zero floats

Custodian Integrations

Custodian	Integration	Auth	Data Retrieved	Sync
CAMS (MF)	API + CAS Parser	OAuth 2.0 + PAN verification	Holdings, transactions, SIP details, dividends	Daily 6 AM IST
KFintech (MF)	API + CAS Parser	OAuth 2.0 + PAN verification	Holdings, transactions, SIP details	Daily 6 AM IST
NSDL (Demat)	CDAS API	Digital signature + 2FA	Equity holdings, corporate actions, IPO allotments	Real-time webhook
CDSL (Demat)	easi/easiest API	2FA + OTP	Equity holdings, transactions	Real-time webhook
Account Aggregator	AA API (ReBIT spec)	Consent artifact (FI-based)	Bank statements, FDs, insurance, NPS	On consent, then scheduled

Ingestion Pipeline

Raw Sources

CAMS, KFintech, NSDL, CDSL, AA framework, manual uploads

→

Normalizer

Map custodian-specific formats to unified Intelligence Graph schema. All amounts to paisa (BIGINT).

→

Deduplicator

Same ISIN held at CAMS and KFintech? Merge. Cross-reference by PAN + ISIN + date.

→

Validator

Paisa check: all BIGINT, no floats. NAV cross-check vs AMFI published values. Units > 0. Dates sane.

→

Enricher

Add ISIN/AMFI code, sector, benchmark, geography, asset class. Tag corporate actions. Link to client goals.

→

Intelligence Graph
Write to holdings, transactions, custodian_accounts. Trigger materialized view refresh for AUM rollups.

DLQFailed records routed to Dead Letter Queue for manual review. Alerts RM if client-impacting. Auto-retry with exponential backoff (max 3 attempts). Client sees last-known-good values until resolved.

Market Data Feeds

Source	Data	Frequency	Protocol
NSE (via Data Vendor)	Real-time equity prices, indices	Every 1s during market hours	WebSocket
AMFI NAV API	Mutual fund NAVs	Daily EOD (9 PM IST)	REST (amfiindia.com)
RBI Reference Rates	Currency rates, SGB prices	Daily	REST
BSE	Bond prices, debt instruments	Every 15s	REST API
CRISIL / ICRA	Credit ratings, bond yields	On change	Webhook

The AI Layer

Not a chatbot. Not RAG. Full context window reasoning on the Intelligence Graph.

The AI layer reads the entire Intelligence Graph for a client — all three signal types, full history, RM notes, goals — and produces context. Not instructions. Not chat responses. Context that makes the RM's natural judgment produce the right action.

Models & Roles

Model	Role in the thesis	Input	Latency
Claude Opus 3	Deep synthesis — morning briefings, meeting prep, Council Mode (5-perspective analysis)	Full client Intelligence Graph	<8s
Claude Sonnet 3.5	Context synthesis — pattern matching, action generation, weekly digests	Signal array + RM notes + holdings	<3s
Claude Haiku 3.5	Real-time classification — every client action classified as it happens	Individual signal event	<500ms
text-embedding-3-small	RM note similarity search — “when did we last discuss this?”	Note text	<200ms

Signal Detection 8 functions

Function	What it detects	Model	Trigger
`classifySignal()`	Type, severity, pattern match from raw telemetry	Haiku	Every client action
`detectAnxietyCluster()`	Repeated holding checks in short windows — signals stress before the client calls	Sonnet	Hourly batch
`detectLoginAnomaly()`	2 AM logins, unusual devices, location changes	Rule-based	Every login
`classifySearchIntent()`	What the client is worried about from their search query	Haiku	Every search
`detectGoalDrift()`	Transactions that move away from the investment mandate	Rule-based + Haiku	Every transaction
`aggregateSignals()`	Cluster all signals per client into a priority score	Sonnet	6 AM daily
`detectExportPattern()`	Unusual PDF exports — often precedes RM switch discussions	Haiku	Every export
`scoreEngagement()`	30-day trend — rising engagement vs risk of churn	Rule-based	Daily

Context Synthesis 10 functions

Function	What it produces	Model	Trigger
`generateMorningBrief()`	Prioritized briefing for all clients — who needs a call today and why	Opus	6 AM daily
`generateClientContext()`	Full Intelligence Graph synthesis for one client on demand	Sonnet	On demand
`generateMeetingPrep()`	Agenda, talking points, personality profile, risk flags — 2h before meeting	Opus	Calendar trigger
`synthesizeCouncilMode()`	5 parallel perspectives (Portfolio, Behavioral, Tax, Risk, Relationship) + synthesis	Opus	On demand
`matchHistoricalPattern()`	“Last time this happened (Sept 2024)” — the sentence that changes RM behavior	Sonnet + Vector	With every synthesis
`generateActionItems()`	Specific actions with owner, priority, deadline — never generic advice	Sonnet	With every synthesis
`generateThinkingChain()`	Full transparency: which signals went in, what patterns were matched, why	Sonnet	With every synthesis
`scoreConfidence()`	Confidence % on every synthesis output — RM sees the uncertainty	Rule-based	With every synthesis
`generateWeeklyDigest()`	7-day summary for RM — what changed, what to watch	Sonnet	Weekly
`summarizeForClient()`	Strip RM-only context, produce client-safe version of the same analysis	Haiku	On demand

Financial Calculations 12 functions

Function	What it calculates	Implementation
`calculateXIRR()`	Annualized return on irregular cashflows, SIP-aware	Newton-Raphson (in-house)
`calculatePortfolioXIRR()`	Weighted XIRR across all holdings, multi-custodian unified	In-house
`calculateLTCG()`	Long-term capital gains with Budget 2025-26 rules and grandfathering	In-house
`calculateSTCG()`	Short-term capital gains at 20% rate	In-house
`calculateDebtTax()`	Debt fund tax at slab rate, no indexation post-2023	In-house
`simulateTaxHarvest()`	Sell + rebuy vs hold: net tax savings with all costs factored	In-house + Haiku
`projectGoal()`	Goal projection with probability distribution — 10,000 Monte Carlo runs	Monte Carlo (in-house)
`calculateAttribution()`	Why did the portfolio return X? Brinson-Fachler decomposition.	In-house
`reconcileMultiCustodian()`	Same ISIN across custodians — unify, flag conflicts, surface errors	Rule-based + Haiku
`calculateDayChange()`	Today's P&L across all holdings in paisa	In-house
`formatIndianCurrency()`	BIGINT paisa → “₹12,84,700” — Indian locale, lakh/crore grouping	Pure function
`calculateRebalance()`	Trades to reach target allocation with tax impact and tracking error	Constraint optimization

Supporting Functions 17 functions

Behavioral Tracking (4)

trackPageView()trackHoldingCheck()trackSearchQuery()trackExportRequest()

Communication (6)

encryptMessage()decryptMessage()formatMessageForDisplay()detectFinancialFigures()generateDocumentSummary()trackMessageSentiment()

Security & Compliance (7)

enforceRLS()hashPAN()auditLog()validateSession()checkFATCA()validateInvestmentMandate()detectAMLPattern()

The Dual Interface

Same Intelligence Graph. Two instruments. Each responds to the other's cues.

These are not two products. The client's behavior on their dashboard is intent data the RM never gets. The RM's context about the client is what makes the dashboard feel personally built. One graph. Two interfaces. A flywheel.

Client Dashboard

apps/client · Next.js 15 · App Router

Unified portfolio view across all custodians
Attribution: why did my portfolio return X%?
Tax impact simulation before rebalancing
Document vault with full audit trail
Encrypted messaging (WhatsApp-feel, not support ticket)

Every page view, holding check, search query, and export is captured as a behavioral signal. The client sees a portfolio viewer. The system sees intent data.

←

Client Intelligence
Graph

→

RM Command Center

apps/rm · Next.js 15 · App Router

Morning briefing: synthesized overnight changes across all clients
Priority queue: ranked by signal urgency, not alphabetical
Client deep-dive: full Intelligence Graph for meeting prep
Post-meeting notes: feed directly back into the graph
Council Mode: 5-perspective analysis for complex situations

The RM does not see raw clickstream. They see synthesized context. The graph produces the insight. The RM provides the judgment.

From Client Action to RM Briefing

Client checks healthcare holdings at 2 AM

trackPageView() and trackHoldingCheck() fire silently. Both write to behavioral_signals table with timestamp, context, time-of-day flag.

↓

Signal classified in real-time

classifySignal() via Haiku: type 'holding_check', severity 'elevated'. Time-of-day (2 AM) increases severity score automatically.

↓

3
Anxiety pattern detected
detectAnxietyCluster() on 48h window: 6 checks in 3 hours on healthcare. matchHistoricalPattern() returns Sept 2024 Biocon episode. Confidence: 87%.

↓

Context snapshot generated

generateClientContext() synthesizes: signals + historical match + RM notes about daughter's medical school fees + client's communication style.

↓

RM morning briefing

8 AM: RM sees 'Mr. Sharma checked healthcare holdings 6x (2-5 AM). Pattern matches Sept 2024 Biocon panic sell. Medical school fees funded from these holdings. Responds well to data-backed reassurance.'

Morning Briefing Generation

Cron 5:30 AM

Fetch overnight signals for all clients. Threshold check: clients with signals > 3 proceed to processing pipeline.

↓

Signal processing per client

aggregateSignals() clusters by type. detectAnxietyCluster() checks for stress patterns. scoreEngagement() computes 30-day trend.

↓

3
Priority ranking
Composite priority_score: anxiety 40%, portfolio change 30%, engagement shift 20%, calendar proximity 10%. All clients ranked.

↓

Briefing generation

Top 10 clients sent to generateMorningBrief() via Claude Opus. Full Intelligence Graph context. Output: structured briefing with talking points, risk flags, suggested actions.

↓

Storage + delivery

Stored in context_snapshots (type: morning_briefing). Push notification to RM app. RM opens at 8 AM and sees streaming text render.

Monorepo Structure

APPLICATIONS

apps/client

Client Dashboard

apps/rm

RM Command Center

↓ ↓

DOMAIN PACKAGES

packages/intelligence

Signal Types, Graph Query, Pattern Processing

packages/ai

LLM Orchestration, Prompts, Streaming

packages/finance

XIRR, Tax, Attribution, Monte Carlo

↓

DATA & DESIGN

packages/supabase

Schema, RLS Policies, Typed Client, Edge Functions

packages/ui

Design System, Tokens, Shared Components

packages/shared

Types, Utils, Constants, Formatters

Security Architecture

UHNI clients demand this. Six layers from network edge to AI safety.

Layer 1Network

Vercel Edge Network — global DDoS protection, 300+ PoPs
HTTPS everywhere with HSTS preload header
Rate limiting: 100 req/min per client, 500 req/min per RM
Geo-restricted to India (configurable per deployment)

Layer 2Authentication

Client: Magic link (email) — no passwords for UHNI. Google OAuth secondary.
RM: SAML SSO via firm identity provider + OTP fallback
Session: Supabase Auth (JWT, 1hr expiry, automatic refresh)
Invite-only: no self-registration — clients onboarded by their RM

Layer 3Authorization

Row Level Security (RLS) on ALL 13 tables — no exceptions
Client role: sees own data only (client_id = auth.uid())
RM role: sees assigned clients only (rm_id = auth.uid())
RM Admin role: sees team clients (team matching)
No cross-tenant data leakage possible at database level

Layer 4Data Protection

PAN: SHA-256 hashed with per-client salt — never stored raw
Messages: AES-256-GCM end-to-end encryption — server cannot read
Documents: encrypted at rest (AES-256) via Supabase Storage
Account numbers: one-way hashed — displayed as masked (****1234)
All money: BIGINT paisa — no floating point errors, ever

Layer 5Audit & Compliance

Append-only audit log: every action recorded, no UPDATE or DELETE
IP address + user agent + timestamp on every log entry
RBI data residency: all data in India-region Supabase instance
DPDP Act 2023: consent management, data minimization, right to erasure
SEBI advisory: investment mandate validation on every transaction

Layer 6AI Safety

No PII in LLM prompts: client names anonymized before sending to Claude
No financial advice in AI output: context and analysis only, never 'buy' or 'sell'
Confidence scoring on all AI outputs: RM sees the % and contributing factors
Human-in-the-loop: RM reviews and approves all AI-generated action items
Thinking chain: full transparency on how every briefing was generated

Database Schema

13 tables. Every one with RLS. All money in paisa. The novel tables are highlighted.

The three Intelligence Graph tables — behavioral_signals, rm_context_notes, and context_snapshots — are what make this system different. Every other wealth platform has the first six tables. Only Jugalbandi has the last three.

Intelligence Graph Tables the novel ones

BEHAVIORAL_SIGNALS

signal_id — PK, UUID, gen_random_uuid()

client_id — FK → clients

signal_type — ENUM (page_view, holding_check, document_access, message_sent, login_pattern, search_query, export_request, feature_interaction)

signal_data — JSONB (event payload)

source — ENUM (client_dashboard, rm_input, system_detected)

created_at — TIMESTAMPTZ DEFAULT now()

expires_at — TIMESTAMPTZ (TTL for raw telemetry)

RLS: Client: NONE — telemetry is not exposed to client. RM: assigned client signals. System: write-only.

RM_CONTEXT_NOTES

note_id — PK, UUID

rm_id — FK → relationship_managers

client_id — FK → clients

note_type — ENUM (general, meeting, call, observation, strategy)

content — TEXT (encrypted at rest)

tags — TEXT[]

linked_holdings — UUID[]

linked_goals — UUID[]

is_meeting_note — BOOLEAN DEFAULT false

meeting_date — DATE

sentiment_score — NUMERIC(3,2) (AI-generated)

created_at — TIMESTAMPTZ

RLS: Client: filtered view (no RM-only tags). RM: own notes for assigned clients only.

CONTEXT_SNAPSHOTS

snapshot_id — PK, UUID

client_id — FK → clients

rm_id — FK → relationship_managers

snapshot_type — ENUM (morning_briefing, meeting_prep, alert, quarterly_review)

signals_considered — JSONB (audit: exactly what went into this)

context_output — TEXT (the synthesized briefing — plain language)

confidence — NUMERIC(5,2)

priority_score — INT (0–100)

created_at — TIMESTAMPTZ

read_at — TIMESTAMPTZ

RLS: Client: NONE. RM: own snapshots for assigned clients only.

Core Tables every platform has these

CLIENTS

client_id — PK, UUID

rm_id — FK → relationship_managers

family_office_id — FK? → family_offices

full_name — TEXT NOT NULL

pan_hash — TEXT (SHA-256 with salt, never raw)

risk_profile — ENUM (conservative, moderate, aggressive, custom)

investment_mandate — JSONB

kyc_status — ENUM (pending, verified, expired)

aml_status — ENUM (clear, flagged, under_review)

fatca_status — ENUM (compliant, non_compliant, pending)

total_aum_paisa — BIGINT (materialized view)

relationship_start_date — DATE

created_at — TIMESTAMPTZ

RLS: Client: own row only. RM: assigned clients. RM Admin: team clients.

HOLDINGS

holding_id — PK, UUID

client_id — FK → clients

account_id — FK → custodian_accounts

asset_type — ENUM (mutual_fund, stock, bond, fd, ppf, nps, pms, aif, real_estate, gold, unlisted, other)

asset_code — TEXT (AMFI/ISIN/custom)

asset_name — TEXT

units — NUMERIC(20,6)

avg_buy_price_paisa — BIGINT

current_nav_paisa — BIGINT

invested_paisa — BIGINT

current_value_paisa — BIGINT

xirr — NUMERIC(8,4)

geography — TEXT

currency — TEXT DEFAULT 'INR'

updated_at — TIMESTAMPTZ

RLS: Client: own holdings. RM: assigned client holdings.

TRANSACTIONS

transaction_id — PK, UUID

client_id — FK → clients

account_id — FK → custodian_accounts

holding_id — FK → holdings

amount_paisa — BIGINT NOT NULL

direction — ENUM (buy, sell, dividend, switch_in, switch_out, sip)

transaction_date — DATE NOT NULL

settlement_date — DATE

corporate_action_ref — TEXT (nullable)

created_at — TIMESTAMPTZ

RLS: Client: own transactions. RM: assigned client transactions.

CLIENT_GOALS

goal_id — PK, UUID

client_id — FK → clients

title — TEXT NOT NULL

target_paisa — BIGINT

current_paisa — BIGINT (materialized)

target_date — DATE

linked_holdings — UUID[]

strategy_notes — TEXT

priority — INT (1–5)

status — ENUM (active, achieved, abandoned, paused)

created_at — TIMESTAMPTZ

RLS: Client: own goals. RM: assigned client goals.

Communication & Audit Tables

MESSAGES

message_id — PK, UUID

thread_id — UUID NOT NULL (per client-RM pair)

sender_type — ENUM (client, rm)

sender_id — UUID

content_encrypted — BYTEA (AES-256-GCM, E2E)

attachments — JSONB[]

read_at — TIMESTAMPTZ

created_at — TIMESTAMPTZ DEFAULT now()

RLS: Client: own threads. RM: threads with assigned clients. E2E encrypted — server cannot read.

DOCUMENT_VAULT

doc_id — PK, UUID

client_id — FK → clients

uploaded_by — UUID

doc_type — ENUM (cas_statement, tax_report, agreement, kyc, other)

doc_name — TEXT

storage_path — TEXT (Supabase Storage)

checksum — TEXT (SHA-256)

is_signed — BOOLEAN DEFAULT false

expiry_date — DATE

audit_trail — JSONB

created_at — TIMESTAMPTZ

RLS: Client: own documents. RM: assigned client documents.

AUDIT_LOG

log_id — PK, UUID

actor_type — ENUM (client, rm, system, cron)

actor_id — UUID

action — TEXT NOT NULL

entity_type — TEXT

entity_id — UUID

payload — JSONB

ip_address — INET

created_at — TIMESTAMPTZ DEFAULT now()

RLS: APPEND-ONLY. No UPDATE, no DELETE — ever. RM Admin: read team logs. Compliance: read all.

Key Relationships

clients.rm_id → relationship_managers.rm_id

custodian_accounts.client_id → clients.client_id

holdings.account_id → custodian_accounts.account_id

transactions.holding_id → holdings.holding_id

behavioral_signals.client_id → clients.client_id

rm_context_notes.rm_id → relationship_managers.rm_id

context_snapshots.client_id → clients.client_id

client_goals.client_id → clients.client_id

The Flywheel

Why this gets better with every client interaction

The Intelligence Graph is not static. Every interaction enriches it. Better data produces better context. Better context produces better RM behavior. Better RM behavior produces better client experience. Better client experience produces more engagement. More engagement produces more behavioral signals. The loop accelerates.

Client uses the dashboard

Views portfolio, checks a holding, searches for something, exports a PDF. Normal usage.

↓

Behavioral signals captured

trackPageView, trackHoldingCheck, trackSearchQuery fire silently. Signals stored with full context.

↓

3
Intelligence Graph updated
Behavioral signals join hard data (custodians) and RM context (notes). The per-client model grows richer.

↓

AI detects patterns

detectAnxietyCluster, matchHistoricalPattern, scoreEngagement run on the graph. Patterns surface that humans would miss.

↓

5
RM receives context briefing
Not instructions. Context. The RM's judgment + synthesized context = better action than either alone.

↓

RM acts with better insight

Calls at the right time. Says the right thing. Anticipates needs. Post-meeting notes feed back into the graph.

↓

Client gets better service

Feels understood. Trusts the RM more. Uses the dashboard more. The cycle accelerates.

The Business Case

30–40

Clients per RM today

→

80–90

Clients per RM with Jugalbandi

Avg AUM ~50Cr per client. Each RM manages ~4,250Cr instead of ~2,000Cr. Revenue per RM more than doubles. Client NPS stays flat or improves.