Sumit K.C.
Start a project

Case studies / verified systems

Case studies that leave a trace.

Four shipped systems, read as working instruments: the constraint, the path through it, and the evidence that survived delivery.

Verified delivery traces live archive

Select a trace

01 / verified studyE-commerce

Falcon CRO / queue system

Falcon CRO

A Shopify conversion-rate optimization and behavioral analytics application designed for reliable event processing and strong storefront performance.

Client
Falcon CRO
Industry
E-commerce
Engagement
Shopify conversion-rate optimization and behavioral analytics application
Stack
  • Shopify
  • BullMQ
  • PM2
Illustrative interface / Sample data processing
queue queuedone active jobsafe delivery
payment.capturedsample event / idempotentdone
session.replayedsample event / idempotentdone
cart.updatedsample event / waitingdone
bullmqpm2 clusterretry safe

01 / constraint

The problem had a shape.

The application needed to process events and payments asynchronously and maintain dependable runtime performance for Shopify App Store submission.

02 / solution path

The system answered in steps.

  1. 01

    Asynchronous processing

    Implemented BullMQ processing for background event and payment work.

  2. 02

    Idempotent handling

    Added idempotent event and payment handling to keep repeated deliveries safe.

  3. 03

    Cluster scaling

    Used PM2 cluster scaling to support the application runtime under load.

Observed outcomes

3xP90 latency

Improvement in P90 latency.

8xCore Web Vitals

Improvement in Core Web Vitals for Shopify App Store submission.

Back to index
02 / verified studySalesforce analytics

SFMC Studio / pipeline system

SFMC Studio

A Salesforce analytics platform for ingesting, storing, and aggregating reporting data safely.

Client
SFMC Studio
Industry
Salesforce analytics
Engagement
Salesforce analytics platform
Stack
  • Salesforce
  • MongoDB
Illustrative interface / Sample data stream stable
sourceSalesforcecloud records
bufferQueuesafe delivery
storeMongoDBreporting data
records / batchsample batch
aggregationrestructured
delivery modeidempotent

01 / constraint

The problem had a shape.

Salesforce Cloud ingestion and downstream MongoDB delivery needed a queue-based flow that could preserve safe delivery while improving analytics quality and web performance.

02 / solution path

The system answered in steps.

  1. 01

    Queue-based ingestion

    Structured Salesforce Cloud ingestion around queued processing.

  2. 02

    Safe data delivery

    Implemented safe delivery of ingested records into MongoDB.

  3. 03

    Aggregation restructuring

    Restructured aggregations to improve metrics accuracy and web vitals.

Observed outcomes

5xMetrics accuracy

Improvement in metrics accuracy.

2xWeb vitals

Improvement in web vitals.

Back to index
03 / verified studyKnowledge management

Lexilift AI / rag system

Lexilift AI

A multi-domain RAG SaaS knowledge base that turns uploaded documents into searchable context for streaming answers.

Client
Lexilift AI
Industry
Knowledge management
Engagement
Multi-domain RAG SaaS knowledge base
Stack
  • RAG
  • Embeddings
  • Vector retrieval
  • Streaming API
Illustrative interface / Sample data query ready
sourceproduct-notes.mddocument loaded
chunks
sample context blocks
vectors
embedding map
answer stream

Relevant context selected. The response is assembled from the indexed source and streamed as it becomes available

01 / constraint

The problem had a shape.

The product needed an end-to-end path from source documents to useful, domain-aware context without attaching unsupported numerical outcomes.

02 / solution path

The system answered in steps.

  1. 01

    Document ingestion

    Built the ingestion flow for bringing source documents into the knowledge base.

  2. 02

    Chunking and embeddings

    Split documents into retrieval-ready chunks and generated embeddings for semantic search.

  3. 03

    Retrieval and streaming queries

    Retrieved relevant context, assembled prompts, and exposed a streaming query API.

No numeric outcome is attached to this study. The system flow is documented instead of overstated.

Back to index
04 / verified studyMortgage lending

American Prime Lending / crm system

American Prime Lending CRM

A custom AI-powered lending CRM for lead management, property enrichment, communications, and lending-team workflows.

Client
American Prime Lending
Industry
Mortgage lending
Engagement
Custom AI-powered lending CRM - Irvine, California
Stack
  • GPT-4o
  • NestJS
  • Aircall
  • Twilio
  • Zillow API
  • PostgreSQL
  • Azure
Illustrative interface / Sample data synced
lead / redactedSample leadfollow-up window / todayAI summary ready
activity feed
Aircall / call logged recent
Twilio / message sent recent
Zillow / property enriched recent
webhook queueclearcalendar sync / Google + Outlook

01 / constraint

The problem had a shape.

The lending workflow required a connected CRM with AI-assisted lead management, third-party integrations, reliable webhooks, and synchronized calendars.

02 / solution path

The system answered in steps.

  1. 01

    Full-stack CRM ownership

    Started with frontend delivery and grew into full-stack ownership of the AI-assisted lead-management workflow.

  2. 02

    Connected lending workflows

    Integrated Aircall, Twilio, Zillow API, Google Calendar, and Outlook synchronization into the CRM.

  3. 03

    Reliable queued processing

    Used idempotent, queue-based webhook processing and improved bulk CSV parsing for operational data flows.

Observed outcomes

4xWebhook success rate

Improvement through idempotency and queue-based processing.

5xBulk CSV parsing

Efficiency improvement for bulk CSV parsing.

20+Defects resolved

Critical- and medium-priority defects resolved.

Back to index

Privacy choices

Necessary storage keeps this site working. Optional analytics helps us understand visits and stays off until you choose it.