Selected Work
Control plane for agent harnesses — enables teams to define their own harness without reimplementing the infrastructure behind it.
- → No vendor lock-in — bring your harness, your framework, your model; keys flow as Capability and are never read from os.environ, with framework adapters for Claude Agent SDK, LangGraph, CrewAI, Google ADK, LlamaIndex, or a bare loop
- → Budget control with overflow — per-principal and per-bucket caps with three OverflowPolicy modes (abort, finish_step, finish_run); BudgetGate.check runs before every wire call and the strictest cap wins
- → One OTel trace per run — every node, LLM call, tool call, sandbox exec, checkpoint, and handoff is a span carrying entorin.run_id, entorin.principal_id, tokens, and cost
Python OpenTelemetry Pydantic MCP Claude Agent SDK OpenAI Agents SDK LangGraph CrewAI
Agentic Workflow Orchestrator
Multi-agent orchestration framework built with LangGraph and the Claude Agent SDK that turns a user idea into an implemented sprint via an evaluator-gated revision loop.
- → Designed a Planner → Evaluator → Advisor → Generator loop with revision budgets and advisor escalation on repeated failures
- → Added parallel planner branches seeded with distinct personas and LLM-based selection of the strongest plan
- → Supported Anthropic and OpenRouter backends with per-role model overrides, human-in-the-loop checkpoints, and resumable runs via artifact snapshots
Python LangGraph Claude Agent SDK Anthropic OpenRouter
End-to-End ML Pipeline
MLOps pipeline for income classification, covering data ingestion, model training, serving, and observability.
- → Trained and compared Logistic Regression and XGBoost with full MLflow experiment tracking
- → Served predictions via FastAPI with Prometheus metrics and Grafana dashboards
- → Containerized full stack with Docker Compose — inference, MLflow, Prometheus, and Grafana
MLOps Python FastAPI MLflow Docker Prometheus Grafana
Travel Journal Android App
Led a team of 5 to build an Android app that lets users document travels and generate customizable single-page journals — won honorable mention at the 2021 Allring Competition.
- → Deployed API layers on Firebase and Google Cloud Functions with a Chrome Extension for bookmarking
- → Built web crawlers and NLP modules to extract scenic spots from bookmarked blog articles
- → Won honorable mention at the 2021 Allring Competition
Android Kotlin Python GCP Firebase NLP
Speech Income & Expense Tracker
Led a team of 4 to build a Python desktop app that extracts items and money amounts from spoken sentences using a custom deep learning model.
- → Designed a deep learning model combining PyTorch and Keras for spoken financial entity extraction
- → Conducted independent research on NLP and speech processing literature to inform model design
- → Received the highest score from peer review, presented as a poster
Python PyTorch Keras Deep Learning NLP Speech
News Article Simplification for Children
Built a system that simplifies news articles for young readers by replacing difficult words with simpler counterparts using BERT and word embeddings.
- → Built a child-oriented lexical resource achieving 75% vocabulary coverage across evaluated articles
- → Preserved semantic fidelity while reducing overall article complexity
- → Improved readability and accessibility for elementary-level readers
Python BERT Word Embeddings NLP
Food Waste Reduction Platform
Collaborated with a team of 5 to build a full-stack web application connecting food donors with recipients to reduce waste.
- → Led development of the API Gateway and Subscription Service, delivered within 2 months
- → Managed data storage with Firebase and containerization with Docker
- → Built on a Service-Oriented Architecture with React and FastAPI
React JavaScript TailwindCSS FastAPI Firebase Docker