Engineering Whitepapers

Technical Architecture & Production Case Studies

We don't publish vague marketing fluff. Here are the exact system architectures, prompt pipelines, guardrail layers, and database schemas behind our live production deployments.

31%

Patient No-Show Reduction (DocBooking)

52%

Less Front-Desk Staff Time (DocBooking)

72%

Faster Lead Qualification (Expreality)

< 2s

P95 Inference & Webhook Latency

Healthcare SaaS ArchitectureProduction Live · Panipat Clinics

Case Study 1: DocBooking Clinical Intake & Triage Engine

An end-to-end multi-agent clinical intake system operating on WhatsApp Cloud API and Next.js, serving private clinics and hospital practices with automated symptom triage, doctor slot allocation, and reminder loops.

System Architecture Diagram (DocBooking Flow)LangGraph + PostgreSQL
[ Patient WhatsApp / Web UI ]
            │
            ▼ (Webhook HMAC-SHA256 Auth)
[ Ingress Fastify Gateway ] ──► [ PII & Redaction Sanitizer (Regex + Presidio) ]
                                          │
                                          ▼
                               [ LangGraph Orchestrator ]
                     ┌────────────────────┴────────────────────┐
                     ▼                                         ▼
         [ Clinical Triage Node ]                   [ Appointment Booking Node ]
        (RAG: Pinecone + ICD-10 SOPs)              (Lock-safe Calendar Slot Matrix)
                     │                                         │
                     ▼                                         ▼
         [ Medical Guardrail Filter ]              [ PostgreSQL Multi-Tenant DB ]
       (Red-flag emergency escalation)             (Prisma ORM, Row-Level Security)
                     │                                         │
                     └────────────────────┬────────────────────┘
                                          ▼
                         [ Twilio / WhatsApp Emitter ]
                        (Sub-2-minute patient delivery)

1. LangGraph State Machine & Safety Guardrails

The conversation operates on a stateful graph built with LangGraph. Each turn preserves session context while preventing hallucinated medical advice:

  • State Schema: Carries patient history, provisional chief complaint, verified insurance tier, and selected doctor slot ID.
  • Emergency Intercept Node: Deterministic regex + zero-shot classification checks for chest pain, acute dyspnea, or severe trauma. Any positive match triggers an immediate hospital emergency escalation prompt with human desk routing.
  • Deterministic Slot Locking: Booking checks are atomic in PostgreSQL using row-level locking (FOR UPDATE SKIP LOCKED) to eliminate double-bookings during concurrent traffic spikes.

2. Prompt Pipeline & Context Injection

System prompts are version-controlled and enforced with strict output schema guarantees:

// Triage System Prompt Template

SYSTEM: You are the clinical intake assistant for [Clinic_Name].

RULES:

1. Never diagnose conditions or prescribe medications.

2. Ask exactly ONE clarifying question per turn to identify primary specialty needed.

3. Output your reasoning in <clinical_thought> tags and patient response in JSON:

{ "specialty": "Dermatology", "urgency": "ROUTINE", "patient_message": "...", "suggested_slots": [...] }

3. PostgreSQL Database Schema

Normalized relational model supporting high-throughput scheduling with full audit logging:

model Appointment { id String @id @default(uuid()) clinicId String @db.VarChar(64) doctorId String @db.VarChar(64) patientPhone String @db.VarChar(20) // E.164 formatted scheduledTime DateTime status Status @default(CONFIRMED) // CONFIRMED, RESCHEDULED, COMPLETED, NO_SHOW intakeSummary Json? // Structured symptoms & chief complaint extracted by LLM whatsappThreadId String? createdAt DateTime @default(now()) updatedAt DateTime @updatedAt @@index([clinicId, scheduledTime]) @@index([patientPhone]) }

Verified Production Impact

31% Reductionin missed patient appointments via 24h & 2h WhatsApp reminders.
52% Time Savedfront-desk staff no longer handle phone bookings manually.
99.8% Uptimezero double-bookings across 14,000+ completed clinic reservations.
Luxury Real Estate IntelligenceProduction Live · Delhi NCR & Gurgaon

Case Study 2: Expreality Spatial Location Scoring & CRM Engine

An AI property intelligence engine connecting private high-net-worth real estate listings with algorithmic neighborhood scoring, multi-agent lead profiling, and automated WhatsApp nurturing.

System Architecture Diagram (Expreality Pipeline)Spatial AI + CrewAI
[ Inbound Real Estate Lead (Portal / Ads / WhatsApp) ]
                         │
                         ▼
             [ Ingestion Webhook Queue ]
                         │
                         ▼
        [ CrewAI Autonomous Multi-Agent Pool ]
    ┌────────────────────┼────────────────────┐
    ▼                    ▼                    ▼
[ Profiling Agent ] [ Matching Agent ] [ Budget Verifier ]
(Analyzes intent)   (Spatial index)    (Capacity score)
    │                    │                    │
    └────────────────────┼────────────────────┘
                         ▼
        [ Geospatial Scoring Engine (PostGIS) ]
        (Transit distance + Civic index + ROI)
                         │
                         ▼
        [ Sync to Enterprise CRM (Zoho / HubSpot) ]
                         │
                         ▼
        [ Private WhatsApp Nurture Journey ]
        (Personalized PDF brochure dispatched in < 2 mins)

1. Spatial Intelligence & Amenity Indexing

Rather than relying on basic text search, Expreality combines PostGIS spatial queries with dense vector embeddings to calculate dynamic livability scores:

  • Isochrone Drive-Time Analysis: Generates 15-minute commute polygons against Delhi NCR arterial expressways (NH-44, Dwarka Expressway, Golf Course Extension).
  • Price Velocity Scoring: Compares quarterly registry records against asking prices to highlight genuine undervalued inventory.

2. Operational Results & Quantitative Proof

72%

Faster Lead Qualification

37 Hrs

Saved Monthly Per Sales Manager

2.1×

Higher Site-Visit Booking Rate

Enterprise AIOps & Multi-Agent OrchestrationLangGraph · Model Context Protocol (MCP) · n8n

Case Study 3: Production AIOps, Telemetry & Token Optimization

How RisonAI Tech maintains enterprise LLM pipelines with OpenTelemetry tracking, semantic caching, prompt drift calibration, and zero unmonitored failures.

Semantic Caching

Embedding-based caching deflects 43% of identical user queries before reaching frontier LLMs, reducing per-turn token costs from $0.04 to $0.003.

Drift Detection

Automated weekly regression runs against synthetic test sets ensure prompt behavior remains deterministic across upstream OpenAI and Anthropic model revisions.

MCP Tool Safety

Model Context Protocol tools run in sandboxed Docker workers with read-only database connections and scoped bearer tokens for write actions.

Verified Third-Party Footprint & Industry Consensus

RisonAI Tech maintains verified external technical profiles and directory citations to eliminate knowledge graph ambiguity:

IndiaMARTVerified B2B CatalogueJustdialVerified Business PanipatGitHubOpen-Source Code ReposClutchVerified Agency Profile