Building intelligent systems from data to deployment.

Agent workflows, data pipelines, backend systems, and customer-facing applications. From Prompt to Production

Signature project

SENTINEL — Trust & Safety Triage Agent

Problem: Unguarded LLM triage produced plausible-sounding enforcement rationales with hallucinated or out-of-scope policy citations—and no automated accuracy regression.

Built: A local LangGraph agent with policy-RAG retrieval (Chroma), code-level citation grounding, layered input/output guardrails, and a CI golden-set eval gate.

Outcome: 0 ungrounded escalations across 75+ live cases; 90% routing accuracy on the CI gate; 0.699 Ragas faithfulness.

Impact

Selected outcomes in production

Backed by measurable outcomes production ownership

Open source

Contributions to other projects

Synced automatically from GitHub — no manual upkeep.

Experience

Data engineering, document pipelines, and applied research.

Graduate work at CU Boulder plus production-style internship delivery—OCR ingestion, Spark ETL.

Jun 2025 — Aug 2025

Data Engineer Intern

Frazier Simplex Machine Company · Washington, PA

Independently architected an OCR-based document ingestion pipeline with Spark and AWS RDS for large-scale CAD drawing digitization.
- Designed ETL workflows and optimized PostgreSQL schema to process 50 GB of CAD engineering drawings.
- Automated digitization of 5,000 drawings—80% faster processing and elimination of manual effort.
- Built real-time Tableau dashboards from transformed data to accelerate planning and vendor selection workflows.
Aug 2024 — May 2026

Master of Science, Data Science

University of Colorado Boulder · Boulder, CO

Graduate study in data science with portfolio work spanning pipelines, and ML systems.

Coursework Machine Learning, Deep Learning and Neural Networks
Jun 2022 — Dec 2022

Machine Learning Engineer Intern

Tech4Good Lab · UC Santa Cruz · Santa Cruz, CA

Built NLP pipelines to cluster and route student feedback for faster, thematic responses at scale.
- Engineered a semantic similarity pipeline using GloVe embeddings on 700 reviews and a cosine-similarity graph between responses.
- Grouped reviews with Spectral Clustering and Louvain methods—0.4 modularity and ~70% less manual review via thematic grouping.

Publications

Transformer-Based Multilingual Automatic Speech Recognition (ASR) Model for Dravidian Languages — Automatic Speech Recognition and Translation for Low Resource Languages (IEEE)
Going Beyond Traditional Methods: Using LSTM Networks to Predict Rainfall in Kerala — Studies in Computational Intelligence (Springer)

Portfolio, in the sky

Beyond the curtain, Apus — the bird of paradise — maps how I build and ship.

Scroll to open the window, approach the constellation, and follow each star to a case study surface.

Approach

I design for what breaks in production—not happy-path demos.

Each case study ties a failure mode to a decision and a measured outcome.

What happens when…

Pick a failure mode to see the response path and where it was proven in production— traffic → Backend & Cloud, tool failure → Agent Benchmarks, model drift → Inference & Evals.

Decision log

Notable ADRs across projects

An ADR documents the tradeoff, the choice, and how we measured impact—each card links to the full case study.