- High-Level System Architecture (End-to-End) flowchart TB U[User (App/Web/Voice)] --> UI[Frontend UI: Chat + Dashboard + Simulations] UI --> AUTH[Auth (Firebase Auth)] AUTH --> API[API Gateway / Backend]
API --> MCP[Fi MCP Server: Secure Financial Data] MCP --> NORM[Data Normalization + Enrichment] NORM --> FS[(Feature Store / Firestore)]
API --> ORCH[AI Orchestrator] ORCH --> LLM[Local LLM: Intent + Reasoning + Explanation] ORCH --> TOOLS[Tool Router]
TOOLS --> SIM[Deterministic Finance Engine (Simulations)] TOOLS --> CAT[Txn Categorization Model (XGBoost/BERT)] TOOLS --> ANOM[Anomaly Models (IsolationForest/Prophet/ChangePoint)] TOOLS --> RISK[Risk Profiling Model] TOOLS --> REC[Personalization Engine (Ranker/Bandit)]
SIM --> OUT[Response Builder: Visuals + Next Actions] CAT --> OUT ANOM --> OUT RISK --> OUT REC --> OUT LLM --> OUT
OUT --> UI
Interpretation: LLM is the “brain” for conversation and reasoning, but calculation + alerts + personalization come from dedicated model services.
- Conversational Query → Intent → Tool Calling (Sequence Diagram)
This is the most important flow.
sequenceDiagram participant User participant UI as Chat UI participant Orchestrator participant Local LLM as Local LLM participant Tools as Tool Router participant Fin as Finance Engine participant FS as Feature Store participant Resp as Response Builder
User->>UI: "Can I afford a ₹50L home loan?" UI->>Orchestrator: Send query + user_id Orchestrator->>FS: Fetch user context (income, spends, loans, net worth) FS-->>Orchestrator: Structured context
Orchestrator->>Local LLM: Prompt(query + context + guardrails) Local LLM-->>Orchestrator: Intent JSON + missing params + tool plan
alt Missing details Orchestrator->>UI: Ask clarifying question User->>UI: tenure/down payment info UI->>Orchestrator: User answers Orchestrator->>Local LLM: Updated context Local LLM-->>Orchestrator: Final tool plan end
Orchestrator->>Tools: Execute simulation tools Tools->>Fin: Loan affordability + cashflow simulation Fin-->>Tools: EMI range + impact + risk score Tools-->>Orchestrator: Computed results
Orchestrator->>Local LLM: Generate explanation + tradeoffs + steps Local LLM-->>Orchestrator: Plain-language response
Orchestrator->>Resp: Build charts + action CTAs Resp-->>UI: Response (text + chart + actions) UI-->>User: Final answer
- Data Pipeline Architecture (MCP → Enrichment → Feature Store → Models)
This diagram shows the machine learning data backbone.
flowchart LR MCP[Fi MCP Data Source] --> ING[Ingestion Service] ING --> CLEAN[Cleaning + De-duplication] CLEAN --> ENR[Enrichment Layer] ENR -->|Merchant normalize| MERCH[Merchant Resolver] ENR -->|Category tag| CATMOD[Txn Categorization Model] ENR -->|Recurring detection| SUBS[Recurring/Subscriptions Detector]
MERCH --> FS[(Feature Store)] CATMOD --> FS SUBS --> FS CLEAN --> FS
FS --> ANOM[Anomaly Engine] FS --> SIM[Simulation Engine] FS --> REC[Personalization Engine] FS --> RISK[Risk Profiling Model]
- Transaction Categorization Model (Hybrid: Rules + ML)
This is how you keep accuracy high in real-world fintech.
flowchart TB TXN[Raw Transaction Record] --> PRE[Preprocessing] PRE --> RULES{Rule-based match?} RULES -->|Yes| OUT1[Category + Merchant (High confidence)] RULES -->|No| EMB[Text Embeddings: Merchant/UPI text] EMB --> ML[XGBoost/BERT Classifier] ML --> OUT2[Predicted Category + Confidence Score] OUT1 --> FINAL[Final Category Output] OUT2 --> FINAL FINAL --> FS[(Feature Store)]
- Anomaly Detection Architecture (Rules + ML + Forecast)
This is how Nivesh generates actionable alerts.
flowchart TB FS[(Feature Store)] --> FEAT[Feature Extraction] FEAT --> RULES[Hard Rules: thresholds (critical)] FEAT --> ISO[Isolation Forest: anomalies] FEAT --> FORE[Forecast Model: Prophet/ARIMA baseline] FEAT --> CPD[Change Point Detection]
RULES --> SCORE[Alert Scoring & Severity] ISO --> SCORE FORE --> SCORE CPD --> SCORE
SCORE --> DEC{Is action needed?} DEC -->|No| LOG[Log only] DEC -->|Yes| ALERT[Generate Alert + Recommendation] ALERT --> GEM[Local LLM: Natural language explanation] GEM --> UI[Push/Inbox Alert in App]
- Simulation & Projection Engine (Deterministic + Monte Carlo)
This ensures trust and removes hallucination risk.
flowchart TB Q[User Question: goal/loan/retirement] --> PARSE[Local LLM -> structured parameters] PARSE --> DATA[Fetch financial context from FS] DATA --> DET[Deterministic Engine: cashflow/EMI/goal math] DET --> MC[Monte Carlo Simulation: return uncertainty] MC --> OUT[Outcomes: P50/P75/P90 + risk bands] OUT --> VIS[Charts Builder] VIS --> EXP[LLM Explanation Layer: assumptions + tradeoffs] EXP --> UI[Final Output]
- Investment Strategy Advisor (Allocation + Rebalancing)
This is strategy-focused, compliant, and scalable.
flowchart TB FS[(User Portfolio + Goals)] --> RISK[Risk Profiling Model] FS --> CURR[Current Allocation Analyzer] RISK --> OPT[Portfolio Optimizer: Risk parity/constraints] CURR --> GAP[Gap Detection: drift vs target] OPT --> REBAL[Rebalancing Plan] GAP --> REBAL REBAL --> EXP[Local LLM: Explain why + risks + alternatives] EXP --> UI[Recommendations + Action buttons]
- Personalization / “Next Best Action” Engine
This is the retention engine.
flowchart TB FS[(User Events + History)] --> CONTEXT[Context Builder] CONTEXT --> RANK[Learning-to-Rank Model (LightGBM/XGB Ranker)] CONTEXT --> BANDIT[Contextual Bandit (Thompson/LinUCB)] RANK --> MIX[Recommendation Mixer] BANDIT --> MIX MIX --> NBA[Next Best Actions: simulations, goals, tips] NBA --> UI[Home feed + prompts + nudges]