Case Studies

Case study 01

Banking & Financial Services

Tier-2 retail bank

increase in release frequency: 26× increase in release frequency
transaction availability sustained: 99.99% transaction availability sustained
core system modifications required: 0 core system modifications required

Timeline: 9 months, delivered in 5 waves
Team size: 11 — architect, 6 engineers, QA lead, SRE, delivery lead, BA

An API layer that let a 22-year-old core banking system ship monthly

The challenge

Every digital feature required a change to a mainframe-era core with a six-month release cycle. The bank's mobile roadmap was effectively hostage to a system nobody was willing to modify.

What we built

We built an API and event layer in front of the core: read models served from a CQRS projection, writes routed through an idempotent command gateway with compensating transactions, and change-data-capture keeping projections current within seconds.

Architecture summary

CQRS with event-sourced projections · Kafka change-data-capture from the core · API gateway with OAuth 2.0 and mutual TLS · Kubernetes on a hardened private landing zone · full audit trail on every command.

Java 21
Spring Boot
Kafka
Debezium
PostgreSQL
Kong
Kubernetes
Terraform

Business impact: Digital feature releases moved from twice a year to fortnightly without a single change to the core system. The core release cycle became irrelevant to the product roadmap.

Case study 02

Healthcare

Multi-site hospital network

reduction in documentation time: 41% reduction in documentation time
coding accuracy vs 91.6% baseline: 94.2% coding accuracy vs 91.6% baseline
clinicians onboarded in 8 weeks: 340 clinicians onboarded in 8 weeks

Timeline: 6 months including a 4-week readiness audit
Team size: 8 — ML lead, 3 engineers, clinical SME, data engineer, QA, delivery lead

Clinical document AI that gave 340 clinicians their evenings back

The challenge

Clinicians were spending an average of 96 minutes per shift on documentation and coding. Two previous vendor pilots had failed on accuracy against real, messy clinical notes.

What we built

A retrieval-grounded documentation assistant trained on the network's own templates and coding standards, with confidence thresholds routing uncertain cases to human review and a clinician sign-off step on every output.

Architecture summary

Claude on a private endpoint inside the hospital's cloud boundary · semantic chunking over clinical templates · hybrid search with clinical-terminology re-ranking · full PHI access audit · human-in-the-loop approval gate.

Python
Claude API
pgvector
FastAPI
HL7 FHIR
Azure
Databricks
Keycloak

Business impact: Documentation time fell by 41% with coding accuracy above the pre-existing human baseline. No PHI left the hospital's cloud tenancy at any point.

Case study 03

Retail & E-commerce

National omnichannel retailer

peak traffic, zero failures: 8.4× peak traffic, zero failures
conversion rate uplift: 22% conversion rate uplift
peak-week losses eliminated: ₹4.2cr peak-week losses eliminated

Timeline: 7 months, live 6 weeks before peak
Team size: 12 — architect, 7 engineers, 2 QA, SRE, delivery lead

A composable replatform that turned peak week from a risk into a record

The challenge

Checkout failures during the previous festive peak cost an estimated ₹4.2 crore in abandoned baskets. The monolithic platform could not scale checkout independently of catalogue browsing.

What we built

A composable rebuild: headless storefront on the edge, checkout extracted as an independently scalable service, real-time inventory orchestration across stores and warehouses, and graceful degradation that sheds non-essential features under load instead of failing.

Architecture summary

Headless commerce with edge-rendered storefront · checkout as an isolated autoscaling service · event-driven inventory orchestration · Redis-backed cart with write-behind persistence · load-shedding by feature priority.

Next.js
Node.js
Kafka
Redis
PostgreSQL
Cloudflare
AWS EKS
Datadog

Business impact: Peak week ran at 8.4× normal traffic with zero checkout failures and a materially better conversion rate than the previous platform's ordinary trading days.

Case study 04

Insurance

General insurance carrier

of claims straight-through: 68% of claims straight-through
faster average settlement: 55% faster average settlement
annualised operating saving: ₹18cr annualised operating saving

Timeline: 8 months across 3 releases
Team size: 10 — architect, 5 engineers, ML engineer, BA, QA, delivery lead

Straight-through processing for 68% of motor claims

The challenge

Motor claims averaged 11 days to settle, with most of that time spent on manual document handling. Adjuster capacity, not claim complexity, was the binding constraint.

What we built

An intelligent document pipeline handling first notification, estimates and supporting evidence, feeding an externalised rules engine that settles low-complexity claims automatically and routes the rest to adjusters with everything already extracted and validated.

Architecture summary

Document ingestion with classification and extraction · confidence-thresholded human review queue · Camunda decision engine owned by the business · event-driven claim state machine · full decision audit for regulator review.

Python
Azure AI Document Intelligence
Camunda
.NET 8
SQL Server
Azure Service Bus
Power BI

Business impact: Settlement time fell from 11 days to 4.9 days overall, with 68% of motor claims settling without adjuster touch. Adjusters were redeployed to complex and disputed claims.

Case study 05

Manufacturing

Automotive components manufacturer

less unplanned downtime: 31% less unplanned downtime
median advance warning: 40 hrs median advance warning
annual downtime cost avoided: ₹2.1cr annual downtime cost avoided

Timeline: 10 months, plant by plant
Team size: 9 — data architect, 2 ML engineers, 3 engineers, OT specialist, QA, delivery lead

Predicting line stoppages 40 hours before they happen

The challenge

Unplanned downtime across four plants was costing roughly ₹2.8 crore a year. Machine telemetry existed but never left the shop floor, and maintenance ran on a fixed calendar regardless of actual condition.

What we built

Edge collectors with store-and-forward buffering feeding a central lakehouse, anomaly and remaining-useful-life models per equipment class, and predictions surfaced directly in the maintenance planning system as schedulable work orders.

Architecture summary

Edge gateways with local buffering for connectivity loss · MQTT to cloud ingestion · Delta Lake time-series store · per-asset-class RUL models with drift monitoring · bidirectional CMMS integration.

Python
MQTT
Databricks
Delta Lake
MLflow
Azure IoT Hub
Grafana
SAP PM

Business impact: Unplanned downtime fell 31% across the four plants in the first full year, with overall equipment effectiveness up 18 points on the worst-performing line.

Case study 06

Logistics & Supply Chain

Third-party logistics provider

fewer status enquiry calls: 61% fewer status enquiry calls
to onboard a new carrier: 3 days to onboard a new carrier
on-time delivery attainment: 92% on-time delivery attainment

Timeline: 6 months to first release, adapters ongoing
Team size: 8 — architect, 4 engineers, data scientist, QA, delivery lead

One tracking view across 47 carriers and four transport modes

The challenge

Customers called for shipment status because no single system knew where anything was. Every new carrier took roughly six weeks of bespoke integration work.

What we built

A canonical shipment event model with carrier adapters conforming to one internal contract, predictive ETA models trained on historical lane performance, and proactive exception alerts issued before a customer notices a delay.

Architecture summary

Canonical event schema with per-carrier adapters · Kafka event backbone with replay · ETA models per lane and mode · customer portal with webhook subscriptions · dead-letter queues with automated replay tooling.

Go
Kafka
PostgreSQL
Python
React
AWS
Terraform
OpenTelemetry

Business impact: Status enquiry calls dropped by 61% and on-time delivery attainment rose to 92%. New carrier onboarding fell from six weeks to three days.

Case study 07

SaaS & Technology

B2B SaaS scale-up

enterprise contracts closed in 2 quarters: 3 enterprise contracts closed in 2 quarters
downtime during tenant migration: 0 downtime during tenant migration
faster customer onboarding: 94% faster customer onboarding

Timeline: 5 months alongside continuous feature delivery
Team size: 7 — architect, 4 engineers, QA, delivery lead

Re-architecting for enterprise buyers without pausing the roadmap

The challenge

The product was winning mid-market deals and losing every enterprise one. Tenant isolation was enforced in application code, there was no SSO, and no usage metering meant pricing was guesswork.

What we built

Progressive migration to database-enforced row-level tenant isolation, an enterprise readiness layer with SAML SSO and SCIM provisioning, usage metering wired to billing, and a documented path to promote large tenants onto dedicated infrastructure.

Architecture summary

PostgreSQL row-level security per tenant · tenant-aware connection pooling · SAML and SCIM identity layer · event-sourced usage metering into billing · per-tenant feature flags and configuration.

NestJS
PostgreSQL
WorkOS
Stripe
Temporal
Kubernetes
Argo CD
PostHog

Business impact: The platform closed its first three enterprise contracts within two quarters of launch. Security questionnaires that previously took weeks are now answered from an evidence library.

Case study 08

Family Office & Wealth

Single-family office

reporting, down from 3 weeks: T+2 reporting, down from 3 weeks
less manual consolidation: 80% less manual consolidation
of figures traceable to source: 100% of figures traceable to source

Timeline: 5 months
Team size: 6 — architect, 3 engineers, data engineer, delivery lead

Quarter-end reporting cut from three weeks to two days

The challenge

Consolidated reporting across eleven custodians, four currencies and a multi-entity trust structure was assembled by hand each quarter, taking three weeks and producing figures nobody could fully trace.

What we built

Automated multi-custodian ingestion with daily reconciliation, an entity and beneficiary model supporting look-through reporting, and a secure portal with per-entity access control and complete audit logging.

Architecture summary

Custodian feed adapters with schema validation · daily automated reconciliation with exception queue · bitemporal position store · entity graph with look-through aggregation · immutable document vault with retention policy.

Python
PostgreSQL
dbt
Airflow
React
AWS
HashiCorp Vault
Snowflake

Business impact: Consolidated reporting moved to a T+2 cycle with every figure traceable to its custodian source. Manual consolidation effort fell by 80%.

Eight engagements, with the numbers our clients agreed to publish.

An API layer that let a 22-year-old core banking system ship monthly

The challenge

What we built

Architecture summary

Clinical document AI that gave 340 clinicians their evenings back

The challenge

What we built

Architecture summary

A composable replatform that turned peak week from a risk into a record

The challenge

What we built

Architecture summary

Straight-through processing for 68% of motor claims

The challenge

What we built

Architecture summary

Predicting line stoppages 40 hours before they happen

The challenge

What we built

Architecture summary

One tracking view across 47 carriers and four transport modes

The challenge

What we built

Architecture summary

Re-architecting for enterprise buyers without pausing the roadmap

The challenge

What we built

Architecture summary

Quarter-end reporting cut from three weeks to two days

The challenge

What we built

Architecture summary

The system you keep postponing is the one costing you most.