Nathan Kim — Statistics & Data Science

I. About

About me

I'm a recent Statistics & Data Science graduate from UC Santa Barbara, with a Technology Management Certification. I'm drawn to transforming complex systems into interpretable insights, revealing patterns that inform meaningful decisions.

Right now I'm building the real-time voice and vision stack for an AI agent that joins Discord calls and plays multiplayer games alongside you, with sub-420ms voice latency and 60+ FPS live game-state detection. I spent the first half of 2026 as a data science intern at BlueAlpha, building synthetic data generators and benchmarking workflows for Marketing Mix Modeling — work I carried into Juno, an agentic copilot that interprets MMM outputs and ships with a benchmark measuring how far its advice can be trusted. Before that I spent nine months at NCEAS designing reproducible pipelines for biodiversity data. On the side, I build agentic AI projects like Romus, a real-time computer vision form coach, and Dialed, a multi-agent guardrail layer for social media.

Outside of code, you will find me reading books, going on runs, exploring nature, or spending time with family and friends.

II. Experience

Where I've worked

AI/ML Engineer Intern @ Stealth AI Startup
San Francisco, CA (Remote)
Jun 2026 — Present
- Built the real-time voice and vision stack for an AI agent that joins Discord voice calls and plays live multiplayer games alongside users, cutting end-to-end voice latency from ~800ms to under 420ms.
- Designed a per-game harness separating Claude-based strategic reasoning from a sub-frame control loop, allowing high-level decisions and frame-level actions to run on independent timing.
- Trained and deployed a 60+ FPS YOLOv8 detector for live game-state extraction, feeding structured observations into the agent's decision layer.
- Implemented Postgres-backed cross-session memory with SQLite failover, plus a Chrome-extension computer-use agent for mid-game web navigation.
Data Science Intern @ BlueAlpha
San Francisco, CA
Jan 2026 — Jun 2026
- Built a synthetic MMM data generator with configurable ground-truth ROI, 5 adstock decay curves, and 3 saturation functions, enabling controlled benchmarking against known parameters.
- Engineered time-series simulation pipelines producing weekly multi-channel spend data with seasonality and stochastic noise across 10+ digital ad platforms.
- Quantified MMM reliability by fitting models to synthetic data and measuring how accurately they recovered known parameters, surfacing systematic bias in high-spend channels under short observation windows.
- Designed stress-testing workflows to evaluate attribution model robustness under varied market conditions, identifying failure modes in how models attribute channel contributions.
Data Science Intern @ National Center for Ecological Analysis and Synthesis
Santa Barbara, CA
Jul 2025 — Mar 2026
- Built reproducible R pipelines transforming CDFW Vegcamp vegetation survey data into VegBank loader format, automating field mapping, cleaning, and normalization across 8 entity types (plots, projects, soils, strata, species, etc.).
- Conducted EDA on 50k+ vegetation records to surface data quality issues and patterns informing downstream spatial and ML modeling work.
- Designed a two-stage transform-and-validate architecture with iterative debug loops, enabling reliable ingestion into VegBank, the national open-access plant community research database.
- Partnered with ecologists and data managers to resolve schema conflicts and align variable formats with federal metadata standards for public dataset release.

III. Projects

Things I've built

A handful of products and experiments. Click any project for the longer story.

Featured

Juno.

2026

0.897 composite100 scenariosLive demo

Agentic copilot for Marketing Mix Model outputs — and a benchmark that measures whether its advice holds up.

A multi-agent system that turns MMM coefficients into prioritized, citation-grounded recommendations — scored against 100 ground-truth scenarios by an LLM judge at 0.897 composite and 0.875 ranking accuracy.

LivePythonFastAPINext.jsClaudeRAGChromaDBSSELLM Eval

Romus.

2026

~30 FPS~200ms latency3 lifts

Real-time computer-vision coach for weightlifting form.

Real-time pose tracking at ~30 FPS with ~200ms feedback latency, paired with a 4-loop agentic AI system delivering personalized voice cues mid-set across 3 compound lifts.

★ BroncoHacks 2026 — Best Use of Backboard

LivePythonFastAPIMediaPipeWebSocketsClaudeRAGComputer VisionAI

Dialed.

2026

5 agentsReal-time inference2 awards

AI guardrails against manipulative social-media content, in real time.

Distributed 5-agent architecture monitoring Instagram in real time, classifying manipulative engagement patterns via an LLM pipeline and intervening before they reach you.

★ BeachHacks 9.0 — Best Mental Health + Best Use of Fetch.ai

LivePythonFastAPIFetch.aiSupabaseWebSocketsBrowser-useElevenLabsAI

Investment Performance Tracker.

2025

ARIMA pipeline2 modes3 error metrics

ARIMA-based forecasting and risk dashboard for portfolios in R Shiny.

A customizable time-series forecasting pipeline with conservative and aggressive modes, plus an interactive Shiny dashboard for risk and return analysis.

LiveRShinyARIMATime SeriesFinance

IV. Connect

Let's get in touch

Best way to reach me is over email. I'm also on GitHub and LinkedIn.

nathank0306@gmail.com

GitHub

@nakim12

in/kim-a-nathan

Nathan Kim.

About me

Where I've worked

AI/ML Engineer Intern @ Stealth AI Startup

Data Science Intern @ BlueAlpha

Data Science Intern @ National Center for Ecological Analysis and Synthesis

Things I've built

Juno.

Romus.

Dialed.

Investment Performance Tracker.

Let's get in touch