Data & AI Case Studies

1The Challenge

Supplier and vendor records from many systems had to be matched to a third-party supplier-data platform in real time. Name-based rules gave too many false matches, and runs that took several hours failed with no way to resume.

2Our Solution

Built about 90% of the platform. It started as an AWS build (Amazon MQ, Lambda, MWAA, ECS Fargate, Bedrock, Redshift) and we led its 2026 rebuild on GCP with RabbitMQ, Cloud Composer, GKE and Snowflake. Matching uses three LLMs (Gemini, Claude, GPT) and only accepts a match when they agree.

3Outcomes

1The Challenge

Every new data source needed custom pipeline code, data quality problems were found too late, and creator-facing analytics were too slow.

2Our Solution

Built a production GCP platform where Airflow DAGs are generated from YAML, so a new source needs config instead of code. Added a custom data-quality framework and moved Creator Studio analytics to ClickHouse.

3Outcomes

1The Challenge

Data from production databases, partner APIs, FTP drops and mobile app events lived in separate places, so the company had no single source of truth for reporting.

2Our Solution

Designed a Snowflake platform with log-based PostgreSQL CDC through Meltano, Dagster orchestration on AKS, a fault-tolerant FTP ingestion service, and a Cube.js semantic layer feeding Looker. Snowflake roles and permissions are managed in Terraform.

3Outcomes

1The Challenge

Engineers were being woken up for pipeline failures that had well-known fixes, such as out-of-memory errors and stuck PostgreSQL replication slots.

2Our Solution

Built a multi-agent system with LangGraph that monitors Dagster, uses an LLM to read failure logs, and applies the known fix: it scales compute and restarts the job, or for replication alerts it stops the job, removes the stuck slot and starts a partial refresh.

3Outcomes

1The Challenge

Analysts needed a single narrative feed covering social and news sources in many languages, but each source had its own format.

2Our Solution

Built a PySpark pipeline on Databricks that pulls Meltwater (X, Facebook, Instagram, Weibo, Douyin, WeChat, news), Reddit, Telegram and Webz into one standard schema, enriches it, validates metrics, and writes Delta tables.

3Outcomes

1The Challenge

Lawyers needed answers grounded in their legal content, not general chatbot responses.

2Our Solution

Built the full RAG pipeline (ingestion, embeddings, Pinecone retrieval, LangChain orchestration) behind a FastAPI service on AWS, with MLflow model versioning.

3Outcomes

1The Challenge

Each new merchant meant rebuilding the same integrations for stores, ad platforms and payments.

2Our Solution

Built custom connectors and reusable ingestion logic for Shopify, Meta Ads, Google Analytics, Uplisting and payment providers. The data feeds automated insights for product and commercial teams.

3Outcomes

1The Challenge

Operational databases, APIs and marketing platforms were not connected, and user-level insights never made it back into the product.

2Our Solution

Combined the sources in BigQuery with Airflow on Cloud Composer, designed ClickHouse tables for fast user-level processing, and built reverse ETL to send results back to downstream applications.

3Outcomes