Kyle Naranjo06 projects · 2022 to Present

Projects

The systems behind the résumé lines: enterprise data platforms, agentic AI, and the research that started it all. Clients are anonymized; the numbers are exact.

Single Customer View

A year-long engagement with a major Philippine bank, partnering with C-suite leaders and business units to ship priority data products on Azure Databricks. The centerpiece is a daily Single Customer View pipeline consolidating ~15 million records from four enterprise systems into ~7 million unique customer keys used across 10+ business units.

Exact matching misses people. A probabilistic record-linkage engine built with Splink surfaced 748,000 candidate duplicate pairs the exact-match logic could not see, scored through a five-tier confidence framework validated at >99.9% accuracy, while pipeline runtime dropped from over 4 hours to under 1 hour.

15M → 7M
records consolidated to unique customer keys
748,000
duplicate pairs surfaced by record linkage
>99.9%
five-tier confidence accuracy
4h → <1h
pipeline runtime

Investment Data Platform

Owned backend feature development across 10+ interconnected microservices powering deal discovery, due diligence, and portfolio monitoring workflows for a large Singaporean investment holding company.

The platform processes terabytes of vendor data from Sustainalytics, PitchBook, Bloomberg, and MSCI through FastAPI services and Dagster pipelines on Kubernetes, landing in Snowflake, with reliability and observability handled through production monitoring on Kibana and Grafana.

10+
interconnected microservices
4
vendor data feeds processed

Enterprise Document Intelligence Platform

Iterated on Snowflake Cortex AI agent workflows with the Knowledge Graph team so investment officers could extract data from long-form documents, run contextual queries, and analyze portfolios at scale.

Owned agent tooling, prompt engineering, evaluation, workflow tuning, and production incident triage across the document-intelligence stack, partnering with a ~15-person cross-functional team spanning frontend, backend, LLM workflows, MLOps, DevOps, and QA.

~15
person cross-functional team

Student-at-Risk Prediction Platform

Spearheaded the infrastructure and ingestion workstream for a student-at-risk prediction platform on Azure Databricks: platform infrastructure provisioned with Terraform, API ingestion built across three source systems, and Excel-based linear-regression models migrated into production Databricks jobs.

Delivery was accelerated by integrating AI coding agents into the workflow with custom agent skills, hooks, and command layers.

3
source systems ingested via API

ChatGPT-to-Databricks MCP Server

A proof-of-concept integration connecting ChatGPT to Databricks through a Model Context Protocol server, the experiment behind the MCP spec's place in the resources list.

The PoC demonstrated a governed path from a conversational assistant to warehouse queries without another one-off API wrapper, and now feeds Kyle's talks on agentic workflows.

AIComprehend

Co-authored and published the IEEE AIComprehend paper, then built and deployed the full-stack Django application used in the study itself.

A four-week controlled study with 58 high school students showed a 13.9% improvement in test scores, research that started the thread running through the rest of this list.

13.9%
test-score improvement
58
students in the controlled study
4 wk
study duration

Have a data or AI problem worth building for?

Get in Touch