KEVIN
JOE
JAISINGH
ABOUT
Computer Science student at Sacramento State. Graduating Spring 2026. I build things with code. Looking for opportunities where I can ship real software.
EXPERIENCE
Building AI solutions for solar and roofing. Owned retrieval pipeline and database architecture. Built RAG system that cut contractor code lookup from hours to under 30 seconds across 487 California municipalities.
Building enterprise AI platforms and ML pipelines. Built a CNN cardiac abnormality classifier achieving 97% AUC-ROC from raw audio recordings.
Managed kitchen operations, staff scheduling, and training. Promoted from team member. Declined return offer to focus on software engineering.
SKILLS
Python, Swift, TypeScript, SQL
FastAPI, Firebase Cloud Functions, PostgreSQL, Docker
React, SwiftUI
PyTorch, LangChain, OpenAI Embeddings, RAG Pipelines
Git, Pytest, GitHub Actions CI/CD, ROS2, Open3D
PROJECTS
Built a complete 3D reconstruction pipeline from scratch using an Intel RealSense D455 depth camera. Instead of relying on out-of-the-box solutions, I wrote custom ICP alignment, loop closure detection, and pose graph optimization — solving real drift and registration problems with math, not magic.
Standard library reconstruction produced noisy, drifted point clouds with ghosting artifacts. Frames accumulated error over time with no correction.
Custom pipeline with IMU-guided initial alignment, point-to-plane ICP, loop closure back to frame 1, and Gauss-Newton pose graph optimization distributing error across all frames. Smart pair selection cut multiway registration from 24hrs to 4hrs.
Python / Open3D / NumPy / SciPy / OpenCV / Intel RealSense SDK
Built a complete indoor 3D mapping pipeline using a Livox Mid-360 LiDAR sensor with LiDAR-Inertial SLAM. Walk through a space holding the sensor and the system produces a dense, clean 3D point cloud of the environment — 3.4 million points from a single bedroom scan.
Raw LiDAR frames suffer from motion blur (sensor moves during each 100ms scan) and naive stacking produces twisted, unusable geometry. The SLAM's internal map is too coarsely voxelized for final output.
Used RKO-LIO to fuse LiDAR + IMU at 200Hz for scan deskewing and pose estimation. Custom export script applies rigid-body transforms (quaternion-to-rotation via SVD) to place each deskewed frame in world coordinates, then voxel downsampling + statistical outlier removal for clean output.
ROS2 Jazzy / RKO-LIO / Python / Open3D / SciPy / Livox SDK2 / Ubuntu 24.04
Built a production RAG system at GoodLeap (solar financing) that automates permit compliance research for roofing contractors. Contractors enter an address and receive a structured permit checklist — replacing hours of manual research with a single query.
Permit requirements vary by city, county, flood zone, and project type. Contractors waste hours calling building departments for information scattered across state codes and municipal ordinances.
Agentic RAG pipeline that resolves an address to its jurisdiction, searches indexed building codes via hybrid retrieval, asks clarifying questions when requirements fork, and outputs a structured permit checklist with code citations.
Python / FastAPI / ChromaDB / PostgreSQL / React / Docker / OpenAI + Anthropic APIs
Native iOS app for a faith-based NGO connecting college campuses across America through organized prayer. Users adopt campuses, commit to prayer hours, share testimonies, and join live gatherings — all coordinated through a real-time Firebase backend.
Campus adoption system, prayer hour commitments with push notifications, activity feeds, emergency prayer alerts, live streaming, AI-generated daily prayer images, admin dashboard with role-based access, and campus map view.
Swift / SwiftUI / Firebase (Auth, Firestore, Cloud Functions, Storage, Messaging) / OpenAI API / Google Sign-In
ML pipeline that classifies cardiac abnormalities from heart sound recordings, achieving 97% AUC-ROC on held-out test data. Built end-to-end from raw audio to trained model.
5,240 recordings resampled to uniform audio, converted to mel spectrograms, fed into a CNN. EfficientNet-B0 transfer learning improved minority class recall from 84% to 92%.
Python / PyTorch / EfficientNet-B0 / Mel Spectrograms / Stratified 80/10/10 Split
CONTACT
Open to opportunities. Reach out.