Jean
Johnson

🤖 Agentic AI

A multi-phase autonomous job search agent built with LangChain and Gemini 2.5 Flash. It parses GitHub READMEs into a vectorized ChromaDB knowledge base of personal projects via RAG, scores incoming job postings against my resume, and uses Tavily AI to autonomously search and extract job descriptions from external career sites. Gmail and Google Drive APIs feed real job alerts and resume versions into the pipeline, PostgreSQL persists scored matches, and a Streamlit dashboard surfaces fit analysis with resume optimization tips for each posting — an automated review queue for the daily grind.

Multi-phase autonomous agent with RAG and Gmail integration

MLOps

An automated benchmarking pipeline that evaluates free-tier LLMs across coding, reasoning, summarization, and RAG tasks — replacing intuition-based model selection with reproducible, auditable data. The core engineering decision is the Chain-of-Thought judge: the judge model must cite a direct quote and identify gaps before scoring accuracy, clarity, and completeness, with every verdict traceable in a reasoning audit panel. The pipeline is instrumented end-to-end — per-model latency tracking revealed a 9× gap — with MLflow tracking, BigQuery storage, and a live leaderboard on GCP Cloud Run.

Automated LLM evaluation pipeline with live GCP leaderboard

🤖 Agentic AI

A fully autonomous job search agent built on a LangGraph StateGraph pipeline that runs daily via GitHub Actions. It fetches live engineering roles from Greenhouse and Ashby ATS APIs across 50+ companies, filters out non-US and irrelevant postings, scores each job against my resume using a multi-provider LLM fallback chain (Gemini → Groq → OpenRouter), and delivers a ranked digest to my inbox every morning. A Next.js dashboard tracks applications, surfaces skill demand trends, and monitors LLM token usage per pipeline run.

Autonomous LangGraph pipeline that scores, ranks, and delivers daily job matches

🧠 AI Dev Tool

A intelligent LeetCode buddy to make you a Leetcode Master. Buddy has a plan with 170+ problems across 3 phases and buddy also helps you track your solutions and patterns covered for easy revision. The project allows you to submit the leetcode solutions, uses Gemini AI to enrich your submission with metadata and commits them directly to GitHub. Built on Netlify serverless functions, the app also classifies each solution by pattern, generate approach summaries, infer time/space complexity, surface key insights and gotchas, and auto-flag problems for revisit based on self-rating.

AI-powered leetcode study tracker with automatic GitHub commits

MLOps

An end-to-end MLOps pipeline that predicts which Arxiv ML papers will gain traction — GitHub stars crossing 100 within 60 days — and serves ranked predictions through a research digest API. A Go service ingests papers daily from Arxiv, Semantic Scholar, and HuggingFace; a monthly Prefect Cloud flow retrains an XGBoost model on a 230K-paper seed dataset with Optuna tuning and MLflow tracking on DagShub; exported models ship to HuggingFace Hub and are scored at query time by a Go GraphQL API on Render. Evidently drift reports publish weekly to GitHub Pages, with DVC versioning and Docker CI/CD.

End-to-end MLOps pipeline predicting which Arxiv papers will trend — daily ingestion, monthly retraining, weekly drift reports

Agentic AI

Built at ASU WiCS × OHack 2026, an open-source client and case management platform for the nonprofits that still run on spreadsheets and paper forms. It covers the full caseworker workflow — client intake, visit logging, scheduling, document storage, reporting — with configurable fields and role-based access control. The AI layer is dynamically orchestrated: photo-to-intake, voice-to-case-notes, multilingual intake, pgvector semantic search, handoff summaries, and funder reports each run on an admin-configurable model chain with priority-ordered fallbacks, managed from a live AI Lab panel.

Open-source case management platform with AI-powered intake, transcription, and semantic search

⚡ Industrial AI

An AI-powered asset monitoring platform for offshore oil & gas, ingesting real-time readings from 96 assets and 175 sensors through a 3-layer pipeline. Layer 1 runs four pure detectors — threshold, Z-score, transmitter divergence, corrosion lifetime — handling 95.7% of readings without an LLM. Layer 2 gathers context via two-stage agentic RAG with PageIndex tree navigation (100% Recall@3 vs 33% for naive vector search). Layer 3 reasons with Azure OpenAI into structured, citation-backed insights delivered to a Next.js dashboard, MS Teams, and a Discord Q&A bot.

3-layer AI detection pipeline monitoring 96 offshore assets across 175 sensors in real time

🏥 Health AI

1st Place, FemTech AZ Hackathon 2026 — selected for production by Sapphire Labs and Golden Seeds. A gamified AI lifestyle companion for women with PCOS, built in 24 hours. Users submit symptoms and optional bloodwork and are classified into one of four PCOS subtypes by a rule-based scoring engine grounded in Rotterdam Criteria. A three-agent pipeline runs the experience: an Extractor reads PDF or image lab reports into structured biomarkers, a Recommender generates a subtype-specific care pathway via a multi-provider LLM chain, and a Coach delivers daily goals, XP, streaks, and badges.

LLM-powered PCOS subtype classifier with gamified daily coaching and bloodwork extraction

Multimodal ML · NLP + Vision

A multimodal sentiment classifier built on the MVSA dataset using cross-modal attention and gated fusion. RoBERTa is fine-tuned as the text branch and ResNet-50 as the vision branch; their feature extractors are then plugged into a fusion head where text-guided visual attention and visual-guided textual attention let each modality query the other, and a learned gate decides how much to trust each. Trained on an AWS g5.xlarge, the fused model reaches 63.67% validation accuracy — beating the text-only (60%) and image-only (54%) unimodal baselines and validating the fusion design.

Val Macro-F1: 0.63 on MVSA dataset

🏆 First Prize · Voxel51 Hackathon

First prize winner at the Voxel51 Visual AI Hackathon. A live posture analysis tool built on MediaPipe Pose Landmarker, tracking 33 body landmarks in real time from a webcam feed through a Gradio UI. Each frame is scored on five ergonomic metrics — torso lean, neck flex, head-forward distance, shoulder symmetry, and ear offset from the shoulder centerline — and OpenCV annotates the live video with angles, alignment lines, and actionable feedback the moment posture degrades. A practical desk-health companion that nags you about slouching before your back does.

Won first prize at Voxel51 Hackathon

Computer Vision · Object Detection

Fine-tuned YOLOv8 on the UEC Food-100 dataset — 100 food categories with bounding boxes — to detect dishes in real-world photos, reaching mAP@0.5 = 0.659. The MLOps focus is the deployment path: the trained PyTorch model is exported to ONNX and compiled to a TensorRT FP16 engine, then benchmarked head-to-head against the ONNX baseline to quantify the latency gains of half-precision inference on GPU. Integrated with the Ultralytics workflow for serving predictions, the project covers the full journey from dataset processing and training to an optimized inference engine.

mAP@0.5 = 0.659 · TensorRT FP16 optimized

AI Dev Tool

An MCP server that helps coding agents — Claude Code, GitHub Copilot, Codex — optimize Linux workloads on Arm64. It exposes three offline, structured tools: parsing perf report output into a ranked list of hot symbols, recommending NEON SIMD intrinsics for hot loops via hybrid semantic and keyword search over 110 curated intrinsics, and auditing requirements.txt, pyproject.toml, or Dockerfiles for packages lacking arm64 wheels. Ships as a Docker container with a curated knowledge base and evaluation harness, so agents give grounded Arm-specific guidance instead of hallucinated intrinsics.

MCP server giving coding agents grounded Arm64 optimization tools — perf analysis, NEON intrinsics, dependency audits

🧠 Deep Learning

A deep learning salary predictor trained on 259,936 certified H1B records from US Department of Labor disclosure data. Raw filings were cleaned into 82 engineered features — rule-based job title categorization, seniority extraction, employer tiering, metro classification, and a log1p transform for the skewed wage distribution. A PyTorch MLP with LayerNorm and GELU converges in ~20 epochs to a $14,897 MAE, performing best in the $75K–$200K range. Error analysis shows it underestimates earners above $200K because the dataset lacks years-of-experience. Served via a live Streamlit app.

$14,897 MAE on 260K US government records

IoT / Edge AI

An edge-to-cloud industrial telemetry platform simulating a data center CRAC cooling fleet on AWS IoT SiteWise and Greengrass V2. A Dockerized edge core validates Modbus sensor data, buffers offline readings in SQLite WAL for replay, and publishes over MQTTS into a four-level SiteWise asset hierarchy with transforms, metrics, and alarms. When an alarm fires, an EventBridge-driven Lambda gathers 30 minutes of telemetry and generates a structured AI diagnosis with Amazon Nova Lite into DynamoDB, with a cooldown to cap LLM spend. The full stack deploys via CDK for ~$10–12/month.

AWS IoT SiteWise + Greengrass V2 telemetry platform with GenAI diagnostics on alarm — ~$10/mo

Computer Vision

A visual AI hackathon project built as a FiftyOne plugin that combines facial landmark detection with illusion image generation. The plugin (@local/agents_world) operates on the LFW faces dataset and supports two operators: annotating facial features (eyes, nose, mouth, face bounding boxes) using MediaPipe, and generating illusion images from a text prompt using diffusion models. Generated illusion samples are tagged and linked back to their source control image with metadata including ground truth identity labels, facial feature coordinates, and the illusion source sample ID.

FiftyOne plugin combining real-time facial landmark detection with AI illusion image generation

MLOps

An AWS-managed MLOps project predicting mortgage rate spread (APR minus APOR) on Arizona HMDA data from 2021–2024. A SageMaker XGBoost model is tuned via Bayesian AMT, versioned in Model Registry, and tracked with managed MLflow. Evidently generates per-feature drift reports and model-quality metrics across yearly vintages; when degradation crosses thresholds, CloudWatch alarms fire actionable SNS emails. A champion/challenger retraining loop promotes new models only when they close the accuracy gap on held-out eval slices. Two serving modes — batch transform for ops scoring and a Lambda-backed serverless endpoint behind API Gateway for a live pricing API — plus static ops and pricing dashboards on GitHub Pages. All infrastructure is Terraform-managed.

End-to-end SageMaker pipeline with drift monitoring and auto-retraining

Cloud / AWS

A four-phase AWS application for real-time face recognition from IoT video streams. Phase 1: a two-tier web/app architecture backed by SimpleDB. Phase 2: elastic EC2 autoscaling on the app tier serving a FaceNet recognition model. Phase 3: a serverless Lambda pipeline — MTCNN face detection feeding a recognition function. Phase 4: edge inference on AWS IoT Greengrass, where devices stream frames over MQTT to a Greengrass Core detection component, detected faces flow through SQS to the FaceNet Lambda, and results return to the device via a response queue.

4-phase distributed face recognition pipeline on AWS — from EC2 autoscaling to IoT Greengrass edge inference

Computer Vision

An image-based calorie estimation system: photograph a meal and get an estimate of its caloric content. Built in Python, the pipeline applies computer vision to identify the food items present in a photo and deep learning models to map recognized foods and portions to calorie estimates. The project explores the practical challenges of food recognition — visually similar dishes, mixed plates, and portion-size ambiguity — as a stepping stone toward effortless diet logging, where pointing a camera at your plate replaces manual food-diary entry.

Calorie estimation from food images using deep learning