JF Jonathan Frias

Open to full-time roles · May 2026

Jonathan Frias

Software engineer building AI-powered products. Recent CS graduate from CSU East Bay. Most recently shipped Python + React tooling at Lawrence Berkeley National Lab and built an Android accessibility app combining on-device computer vision with a fine-tuned LLM for the Walkaware research project.

Portrait of Jonathan Frias
Fremont, CA

01. About

Engineering at the intersection of systems, applied AI, and accessibility.

The work I'm proudest of so far comes from problems where the constraint sits with the user, not the framework — an Android app that has to give a visually impaired pedestrian useful guidance fast enough to act on, scientific tooling that has to be unambiguous during a beamline experiment, an emergency-response prototype designed for someone who can't reach their phone.

I transferred into CSU East Bay carrying a 3.229 cumulative GPA and graduated with a 3.4 cumulative / 3.98 in-major GPA across four semesters of CS coursework. I currently spend most of my time building FIRST DAY — an AI agent that turns any GitHub repo into an interactive tutor with chat-over-code citations, auto-generated diagrams, and AI-graded assessments.

2,800+
training samples curated for the SafeStep fine-tuned LLM
6
NLG metrics in the SafeStep evaluation harness (BLEURT, COMET, BERTScore, SBert, USE, METEOR)
~1,000
lines of Python automation shipped at LBNL across four beamlines
1 / 100
"Best Research" at the MESA-U 2024 hackathon

Toolbelt

Languages
TypeScript · JavaScript · Python · Java · Kotlin · C / C++ · Rust · Assembly
Frontend
React · Next.js · Tailwind · Vite · HTML / CSS
Backend
FastAPI · Actix-web · REST APIs · socket programming
Applied AI / ML
Claude API with prompt caching · OpenAI Fine-tuning · TensorFlow Lite · NLG evaluation · CV pipelines
Mobile
Android (Kotlin + Jetpack Compose) · TFLite on-device · Android TextToSpeech
Infra / Tools
Linux · Docker · Tailscale · Git · SciCat · LabView APIs

02. Experience

Three roles, one consistent story: shipping things that have to work.

  1. Jun – Aug 2025 · Berkeley, CA

    Automation & Software Development Engineer Intern

    Lawrence Berkeley National Laboratory · Advanced Light Source · ASPIRES program

    • Built a React + TypeScript interface (forked from als-computing/sample_metadata_app-frontend) for ALS beamline scientists to configure X-ray spectroscopy samples; authored multi-row selection (shift+click, click-drag, shift+arrow), bulk duplicate/delete, and an Undo system that reverts an entire selection in one action.
    • Wrote ~1,000 lines of Python automation (ScanFileGenerator) that authenticates to SciCat via pyscicat and generates LabView scan files for XAS / XLD / XMCD across 4 beamlines (402, 631, 4.0.2, 6.3.1) with element-specific edges for 11 elements.
    • Replaced hours of per-file manual editing (~25–300 hand-written scan files per experiment at ~1–2 min each) with a sub-minute batch run, eliminating per-file format errors during beamline ops.
    • Selected for ASPIRES Summer 2025; delivered a 1-minute lightning pitch and open poster session.
    • React 18
    • TypeScript
    • Vite
    • Python
    • pyscicat
    • LabView API
    • Docker
  2. Jan – Oct 2025 · Hayward, CA

    Undergraduate Research Assistant

    CAHSI Walkaware · PI Prof. Lynne Grewe · CSU East Bay

    • Built SafeStep, an Android accessibility app (~970 LOC Kotlin) running a 3-stage ML pipeline per frame: on-device TFLite object detection → on-device TFLite severity regression (Low/Moderate/High/Critical) → fine-tuned LLM narrative → Android TextToSpeech.
    • Trained both TFLite models from scratch — hazard detector and severity regressor — deployed fully on-device for offline operation and sub-second latency.
    • Curated a ~2,800-row training dataset (~933 photos × 3 narrative variants as data augmentation) and iterated through fine-tunes of GPT-3.5-turbo and GPT-4o Mini for the narrative head.
    • Built a six-metric NLG evaluation harness (BLEURT, COMET, BERTScore, SBert, USE, METEOR); achieved median BLEURT 0.65 / BERTScore F1 0.60 / COMET 0.75 / SBERT 0.80 and diagnosed three error classes (~25% severity / ~20% directional / ~15% over-optimism).
    • Kotlin
    • Jetpack Compose
    • TensorFlow Lite
    • OpenAI Fine-tuning
    • Python / Colab
    • Android TTS
  3. Aug 2024 – May 2025 · Campbell, CA

    Field Service Engineer

    Murray Window & Door

    • Developed a preventive maintenance schedule that contributed to a 15% reduction in product failures.
    • Documented equipment performance metrics in logs and databases to support analysis and reporting.
    • Participated in equipment installation, trial runs, investigative tests, repairs, and overhauls.

03. Projects

Selected work — each one ships, each one solves something concrete.

FIRST DAY

In progress · Jun 2026

AI codebase onboarding tutor.

An AI agent that ingests any public GitHub repository at a pinned commit and turns it into an interactive tutor — chat-over-code with citations, auto-generated Mermaid architecture diagrams, and AI-graded comprehension assessments that gate progress.

  • Next.js 14
  • TypeScript
  • FastAPI
  • Claude API
  • tree-sitter
  • sqlite-vec

SafeStep

CAHSI Walkaware

Android accessibility app for visually impaired pedestrians.

Three-stage ML pipeline: on-device TFLite hazard detection + severity regression, then a fine-tuned LLM narrative spoken via TextToSpeech. Evaluated against six NLG metrics with median BLEURT 0.65 / BERTScore 0.60 / COMET 0.75.

  • Kotlin
  • TensorFlow Lite
  • OpenAI Fine-tuning
  • Android TTS
View on GitHub

MotoCrashApp

MESA-U 2024 · Best Research

Emergency-response prototype for motorcycle riders.

Tkinter desktop app that monitors for a configurable post-incident countdown; if the rider doesn't dismiss it, the system auto-calls the Google Geolocation API and dispatches a WhatsApp message to the emergency contact containing name, allergies, and live coordinates.

  • Python
  • Tkinter
  • PyWhatKit
  • Google Geolocation API

ScanFileGenerator

LBNL · ASPIRES

X-ray spectroscopy automation at LBNL ALS.

Pulls sample configurations from SciCat and generates LabView-format scan files for XAS / XLD / XMCD across four beamlines with element-specific edge energies. ~1,000 lines of Python.

  • Python
  • pyscicat
  • LabView
View on GitHub

Parkly

Team capstone

Multi-threaded Java parking garage management system.

Client–server architecture: multi-threaded socket server, Singleton + Facade patterns, service-oriented design across auth, payment, ticketing, and space-tracking services, with dual Swing GUIs for customer and employee workflows.

  • Java
  • Swing
  • TCP sockets
  • Design patterns
View on GitHub

Home Security NVR

Home-lab

Self-hosted security camera system on a Linux home server.

Frigate (open-source NVR with real-time object detection) running in Docker, integrated with a network camera and exposed remotely via Tailscale mesh VPN for off-network access. Linux sysadmin, containerization, networking, and open-source integration.

  • Linux
  • Docker
  • Frigate
  • Tailscale
  • MQTT

04. Education

A growth story, not a flat number.

California State University, East Bay

B.S. Computer Science · Graduated May 2026

Cumulative
3.4/4.0
CSUEB in-major
3.98/4.0
Transferred in at
3.229/4.0
Raised cumulative GPA from 3.229 to 3.4 across four semesters by performing at a 3.98 in CSU East Bay coursework — growth under pressure rather than a number that flatters.

Coursework — Software Engineering · Artificial Intelligence · Web Development · Data Structures · Algorithms · Statistics · Linear Algebra · Computer Architecture · Programming Concepts · Software Responsibility · Computer Networks

Affiliations — CAHSI (Computing Alliance of Hispanic-Serving Institutions) REU participant. Research in computer vision under Prof. Lynne Grewe.

05. Contact

Looking for the right team.

Currently seeking full-time SWE / Applied AI roles. The fastest way to reach me is email or LinkedIn — I respond within a day.