Agentic AI Systems · Grounded Control

Multimodal
AI agents
you can trust.

I build agentic AI systems for real production use: multimodal audio intelligence, autonomous creative execution, and grounded control that keep powerful agents precise, steerable, and reliable. Selected for Startup UCLA Summer Accelerator 2026.

Startup UCLA Summer Accelerator 2026 Multimodal AI Agents Agentic AI Systems Autonomous Creative Intelligence Grounded Control Incoming Penn MSE '27 UCLA Stats & Data Science '26

View Agents Get in Touch

Accelerator

Startup UCLA Summer Accelerator 2026

Building

Stealth AI audio startup

Cohort

Grant-backed summer build · Demo Day Sep 10

Scroll

01 — About

What I build.

I'm a Statistics & Data Science student at UCLA finishing my B.S. this June, selected for the Startup UCLA Summer Accelerator 2026, then heading to the University of Pennsylvania for an M.S.E. in Systems Engineering.

The question I care about in AI is not whether a model can generate something impressive. It is whether an agent can understand intent, act across modalities, and stay under grounded control when the work actually matters.

That shows up across my work: grounded generation in FishCapsule, deterministic audio-to-visual systems in SpiroMint, and enterprise ETL pipelines at TTI where messy systems had to become dependable enough to drive operational decisions.

I'm now building as Co-Founder & CTO of a stealth AI startup focused on multimodal agents.

02 — Projects

Things I've built.

Stealth AI Startup

Co-Founder & CTO

Multimodal AI Agents.

Selected · Startup UCLA Summer Accelerator 2026

Live site

Flagship System · Grounded RAG

FishCapsule

FishCapsule is a study system built around controlled generation rather than free-form answers. It ingests syllabi, maps courses, selects slides, and keeps every response grounded in the user's actual materials.

The core pipeline combines schema-first generation with Zod contracts, slide-grounded RAG, hash-based caching, and a validation-and-repair layer. That reduced invalid outputs by 70%, kept citation coverage at 99%, and made the system more trustworthy in practice.

Citation coverage99%

LLM error reduction70%

Showcased · PIC Spotlight Jan 2026

Grounded RAG Schema-First Validation Layer Next.js TypeScript Zod

Company

Internship · Enterprise Data Systems

TTI PLM Analytics

Built multi-source ETL systems across Teamcenter and OpenLM under enterprise constraints, where the data was messy, operationally important, and tied to recurring reporting needs.

The pipeline covered 6.18M+ usage logs and 14,624 user records, surfaced 15% underutilized licenses, projected $34K annual savings, and automated reporting that saved 38+ hours per month.

ETL Teamcenter OpenLM Pandas SQL

Devpost

Live Demo · SpiroMint — Deterministic Provenance Pipeline

Rotors 3

Speed 1.0x

Complexity 4

Trail 0.04

Palette

SpiroMint turns audio signals into deterministic parameter sets and then into parametric animations using multi-rotor spirograph curves. The chain from signal → parameters → output is reproducible from audio metadata, making provenance inspection possible before minting via IPFS.

↑ Use the controls to inspect the pipeline. Each parameter set represents an explicit stage in how signal drives output.

Signal Mapping Deterministic Output Meyda IPFS FFT

View Devpost

🏆 3rd / 113 · AI Best Coast 2025

Blockchain Winner Fintech Winner

Research

Research · Structured Transformation

Music-Driven Generative Research

Undergraduate research with Prof. Jiayin Lu focused on turning audio features into interpretable geometry. Beats, RMS, spectral centroid, and Mel-band peaks are mapped into parametric curves as a structured module for UCLA's Math–Code–Art initiative.

Built an end-to-end data → geometry → pixels → video pipeline using NumPy, PIL RGBA compositing, and FFmpeg.

Signal-to-Structure Interpretable Mapping NumPy PIL FFmpeg

GitHub

Capabilities in Public Work

Grounded RAG + validation

FishCapsule

Multi-source enterprise data systems

TTI

Deterministic signal pipelines

SpiroMint

Signal-to-structure mapping

Research

Full-stack implementation

TypeScript / Python

Kaggle

Secondary · ML Evaluation

Skin Cancer Classification

A smaller benchmark-style project focused on disciplined model evaluation on 50K dermatology rows. Unified preprocessing and 5-fold comparison across Logistic Regression, Random Forest, and XGBoost.

Final ridge-regularized logistic model scored 0.607 AUC and ranked 7th among ~170 students.

Final rankingTop 4%

Test AUC0.607

Model Evaluation Cross-Validation XGBoost Python

03 — Journey

How the systems thinking deepened.

Jul – Aug 2025

Techtronic Industries · First encounter with messy production systems

Worked across Teamcenter and OpenLM, built ETL around 6.18M rows, and learned that real operational value depends on systems staying reliable under imperfect data.

Sep 2025

Undergraduate Research · Representation became a systems question

At UCLA's Math–Code–Art initiative, I focused on turning signals into interpretable structure — a shift toward controllability, not just output novelty.

Nov 2025

SpiroMint · Public proof of deterministic generation

3rd place at AI Best Coast with an audio → parameters → output pipeline where provenance could be inspected and reproduced rather than treated as a black box.

Jan 2026

FishCapsule · Trust moved to the center of the product

Showcased at PIC Spotlight with a schema-first, grounded RAG system built around validation, repair, and citation coverage instead of demo-only generation.

Jun 2026

Graduating UCLA · Stronger judgment around what makes AI usable

B.S. Statistics & Data Science with public work spanning enterprise pipelines, grounded LLM systems, and deterministic generative systems.

Jul 6 - Sep 10, 2026
Selected for Startup UCLA Summer Accelerator 2026
Selected as Co-Founder & CTO of a stealth AI startup focused on multimodal agents.

Aug 2026 →

Incoming MSE · University of Pennsylvania

Systems Engineering as the next step toward deeper work in controllable AI systems, agent reliability, and grounded control.

Best fit

Agents
with judgment.

Best fit for agentic AI, multimodal systems, founding-engineer, and applied AI product roles where intelligent agents need grounded control and production-level reliability.

Start a Conversation GitHub

harleyjia123@gmail.com

@RukaAtre1

SpiroMint Devpost

Los Angeles, CA

→ Philadelphia, PA · Fall 2026

Multimodal AI agents you can trust.

Multimodal
AI agents
you can trust.