Irvine, CaliforniaAI · software · research

About me

Hi, I’m Baelyn.

I’m a builder, gym girlie, food connoisseur, researcher, founder, and creator.

My work spans mortgage AI, language-model reliability, exoplanet classification, and civic software.

$125k+aigents ARR

50,000+Conversations handled

JEIResearch accepted by JEI

Selected work.

01 / COMPANY · APPLIED AI

aigents

From a problem my mom was facing to AI agents used by mortgage professionals at Rocket Mortgage's Largest US Broker.

See Aigents in action Explore other work

$125kAnnual recurring revenue (ARR)

50,000+Conversations handled

10,000+Appointments booked

5,000+Hours of manual work saved

01 / The spark

It started with my mom’s work.

She always came home from the office tired, eyes constantly drifting to her phone, and our precious time together constantly interrupted by calls and texts. I began to wonder...was there a way to automate her manual workload?

Visit website

02 / The system

The system behind the conversation.

I connected event-driven workflows, LLM agents with tool calling, persistent conversation memory, CRM and scheduling APIs, and self-hosted infrastructure. I taught myself what each part needed, using AI to self-teach. My learning strategy was "learn by doing."

02 / System sketchBehind a single conversation

Lead repliesCRM + messaging

Workflow orchestrationEvent routing & follow-up

LLM agentsTool use & qualification

Conversation statePersistent context & history

Calendar toolsAvailability & booking

Loan officerHuman handoff

Simplified architecture · self-hosted services

03 / In action

From first reply to booked.

The moment leads reply, the aigent kicks in (speed matters!). While it answers questions, gathers info, and nurtures the prospect, the CRM stays up to date, tasks move forward, and the loan officer gets all the conversation details.

Alex JordanHELOC inquiry

Alex

We’re finally redoing our kitchen. Thinking about a HELOC.

Lead workspaceUpdated as the conversation unfolds

1Responded

2Qualifying

3Booked

LEAD DETAILSLead received

Purpose: Kitchen renovation
Requested amount: Waiting for reply
Credit score: Not collected yet
Appointment: Not booked yet

TASKS0/3 complete

Follow up with leadPending

Schedule a callPending

Email the loan officerPending

00 / 34s

Select a chapter to explore at your own pace.

The challenge: Keeping conversation context, collected details, and calendar availability consistent from first reply to handoff.

04 / The lesson

What real use taught me.

A good reply is only one part of a useful product. There are so many little things I never thought about until I received real client feedback. How often are follow ups sent? When does human handoff take place? How does the aigent work alongside the loan officer rather than in isolation? In a nutshell? There is so much more to a system than the flashy features you see. It has to be intuitive to use, not just add another headache to deal with.

05 / What’s next

Expanding to full stack mortgage automation.

I’m now extending into Meta lead generation, qualification, and matching higher-intent prospects with the right mortgage programs. We've already launched campaigns for a few clients that are driving cheaper outcomes, more conversion, and higher qualified leads. After mastering that? Our next stop: automating underwriting/document processing. That is the full funnel system.

NOW EXPLORING

Lead generation · qualification · mortgage-program matching

Keep exploring my work

Research & other builds.

RESEARCH / AI RELIABILITY

LLM Hallucination Research

Journal of Emerging InvestigatorsAccepted for publication

Conducted research at ALPS Lab at UT Dallas under Dr. Gopal Gupta, a CS professor and director of the Center for Applied AI and Machine Learning, with a PhD in computer science from UNC Chapel Hill.
Developed an explainable AI framework combining LLM tool use, Prolog validation, and interpretable checks to study scheduling errors across simple, complex, and ambiguous queries.
Reduced errors by nearly 73% in the study’s scheduling tasks.
Paper accepted for publication in the Journal of Emerging Investigators.
Received commendation from scientific reviewers, PhD students, and professors.
Challenge: Models may "hallucinate" when interpreting real world data or requests. For appointment setting, this means calls booked for the wrong times and giving users false info about calendar availability. For other domains...this could be even more consequential.
I learned: Natural-language fluency does not guarantee correct tool use. Explicit rules and interpretable checks give me a way to examine the model’s behavior rather than rely only on how plausible its answer sounds.

MACHINE LEARNING / SPACE

ExoForge AI

NASA Space Apps ChallengeGlobal Nominee

I built a machine-learning web app to classify exoplanet candidates using NASA Kepler, K2, and TESS data.
Brought predictions together with confidence scores, explanations, visualizations, and simulations, making the model’s outputs easier to explore.
Selected as a NASA Space Apps Global Nominee.
Challenge: I wanted to translate space-data into an educational experience for users to explore potential far-away worlds. The design challenge was to make the prediction, confidence, and scientific context understandable all together.
I learned: Presenting a model’s output is a substantial part of building an ML product. Confidence and explanations help users ask better questions about what a prediction means.

ExoForge AI app with mission stages, exoplanet classification inputs, and model controls

Watch ExoForge demo

CIVIC SOFTWARE / PRODUCT

CivicLink

Congressional App Challenge2nd place · California’s 46th District

I built a resident reporting app and a dashboard with analytics for city officials to view and track local issues.
Connected resident submissions to an organized response workflow, with community-service and reward features to encourage participation.
Earned 2nd place in the Congressional App Challenge, California’s 46th District.
Challenge: The app has two audiences: residents who need a simple way to report a problem, and officials who need organized information they can act on. Both aspects of the app had to work seamlessly.
I learned: After a report is submitted, the information needs to remain useful after it reaches the person responsible for responding. Helped me understand app development from two perspectives/user-types.

Watch CivicLink demo

PAID WORK / AI SYSTEMS

OmniFusion AI

Worked one-on-one with the CEO as an AI developer on client automations and agent workflows.
Developed workflows, debugged scripts and API integrations, and improved lead-generation and sales-funnel processes.
Turned client requirements into working systems that fit existing tools and handoffs between people and software.
Challenge: The automation needs to fit existing tools and the people using them. You can't expect people to migrate to brand-new systems and ditch their existing ones. Instead, you have to work around them and tailor-fit each solution you build to each client.
I learned: How to make systems people actually want to use. I applied what I learned from working with real clients at OmniFusion AI directly to the practices in my own agency.
Fun Fact! I cold DM'ed the CEO a selfie video explaining my desire for hands on experience, my qualifications, and how much I was inspired by his work. He interviewed me a week later and brought me on board.

OMNIFUSION · COMPANY-WIDE RESULTS

120+Active clients5,800+Calls booked monthly$35M+Revenue generated3 yearsSpecializing in high-ticket sales

Visit OmniFusion

BEYOND THE BUILDS

Teaching, community & movement.

INTERNSHIP / AI MARKETING

HubSpot

Created and delivered a marketing campaign and assets for a biotech client as an AI marketing intern.
Organized and hosted a community chatbot-building workshop, sharing practical AI skills with peers.

COMMUNITY / LEADERSHIP

Girls Who Code

Served as president of Girls Who Code at Irvine Valley College and recruited 23 new members.
Hosted coding workshops, planned board meetings, coordinated officer responsibilities, and represented the club at events.
Worked with officers and our advisor to plan career workshops, socials, and opportunities to connect with women in STEM.

CONTENT / MOVEMENT

Yoga & Pilates

Filmed and edited my own yoga and Pilates routines for all skill levels on YouTube.
Shared free at-home practices, bringing my love of gentler movement and mindfulness to others.

Visit my YouTube channel

Education.

UNIVERSITY

University of California, Berkeley

Computer Science · Incoming

CS @ UC Berkeley Also admitted to Carnegie Mellon, Duke, and UMich for CS.

A SELF-DIRECTED EDUCATION

I made room to build.

Cabrillo Point Academy

Class of 2026

GPA: 4.56 · SAT: 1530 · Class rank: 2/163 Coursework includes AP Computer Science A, AP Statistics, AP Physics C, and college-level calculus and linear algebra. Dual enrollment: Irvine Valley College, Coastline College, Laney College, and Santa Ana College. I chose to leave public school for independent study, so I could design a personalized curriculum and have room for building. I wanted my education to revolve around my interests and real world endeavors, not be two separate tracks.

LCAP Advisory Board Member — Brainstormed ways to improve our school with fellow board members and policy makers Step Up Tutoring — Volunteer tutor for underserved elementary school students National Honor Society — Helped package food at food banks, raised money for refugees, promoted public transport at OCTA

Honors.

Congressional App Challenge

2nd place · California’s 46th District CivicLink

NASA Space Apps Challenge

Global Nominee ExoForge

CAASPP

Highest attainable scores in Mathematics and English Language Arts · Grade 11

Journal of Emerging Investigators

Accepted for publication An explainable AI framework for studying query complexity and hallucinations in LLM scheduling tasks