Home

About

Journey

Skills

Contact

Hi there! 👋

I'm Parth Koshti.

Based in Pune, India, I'm an AI/ML Student who builds systems across computer vision, geospatial ML, and LLM-powered multi-agent systems focused on shipping things that actually work in production.

Get in Touch!

AI Engineer

PyTorch Python OpenCV FastAPI LangChain AWS Docker CUDA

ABOUT

Hi, I'm Parth, an AI/ML engineer based in Pune, India, pursuing a B.E. in AI & ML at Symbiosis Institute of Technology. I'm currently an AI Engineer Intern at Akiyam Solutions, building LangGraph multi-agent systems for a geospatial SaaS platform.

What draws me to AI is turning messy real-world data like satellite imagery, LiDAR scans, and financial documents into systems that actually reason and act on it. I work across computer vision, geospatial ML, and LLM agent pipelines, with PyTorch, LangGraph, and FastAPI as my go-to stack.

I care about building things that actually work in production, not just in a notebook.

My Journey

Journey Timeline

AI Engineer Intern @ Akiyam Solutions Private Limited

Jan 2026 - Present

Architected a 21-node LangGraph multi-agent mesh (3 supervisor clusters) for a production-grade geospatial SaaS platform handling flood risk, climate projection & CSRD compliance use cases. Integrated 8 physics-based simulation engines with a Fourier Neural Operator surrogate model achieving 100x speedup and sub-2-second p99 latency. Built FAISS/ChromaDB vector search pipelines and deployed 50+ REST/WebSocket endpoints.

Pune, India

B.E. in Artificial Intelligence & Machine Learning @ Symbiosis Institute of Technology

2024 - 2028

Pursuing a Bachelor's in AI & ML with a GPA of 8.09, building a strong foundation across computer vision, deep learning, and data engineering while shipping real-world projects on the side.

Pune, India

Google Campus Ambassador

Symbiosis Institute of Technology

Represented Google on campus, promoting developer programs and technical initiatives to fellow students.

Pune, India

Senior Editor, Department Magazine

Symbiosis Institute of Technology

Led editorial planning, content review, and coordination with contributors for the department-level academic publication.

Pune, India

SKILLS

CV / Vision

PyTorch TorchVision OpenCV YOLOv8 EfficientNet CUDA Grad-CAM Transfer Learning

Geospatial / 3D

Rasterio GDAL Open3D SentinelHub CesiumJS Point Cloud (LiDAR) Satellite Imagery

AI / LLM

LangChain LangGraph RAG Pipelines HuggingFace Transformers FAISS Vector Search

ML / Data

Scikit-learn Supervised Learning EDA NLP Embeddings Statistics Visualization

Backend

FastAPI Django REST APIs WebSocket OOP Modular Architecture

Cloud / DevOps

AWS (S3, EC2) Docker Kubernetes Git GitHub CI/CD Kafka KEDA

PROJECTS

GSA-SIP: Geo-Sim Intelligence Platform

Designed a 21-node LangGraph AI agent mesh routing queries across 3 supervisor clusters and 6 domain agents powering a multi-tenant SaaS with flood risk, wildfire simulation, and TCFD/CSRD compliance workflows. Integrated Delft3D/SFINCS physics engines with an FNO surrogate model delivering 100x speedup with under 2s p99 latency, streaming inundation results to a CesiumJS 3D globe via WebSocket in real-time. Deployed on Kubernetes GPU pods with KEDA autoscaling (2-20 pods) and a Kafka queue.

LangGraph FastAPI PyTorch (FNO) Kubernetes SentinelHub CesiumJS

View on GitHub

Aerial LiDAR Point Cloud Segmentation

Built a 3D semantic segmentation pipeline on the DALES aerial LiDAR dataset, fine-tuning PointNet++ for 8-class point cloud segmentation (buildings, vegetation, vehicles, power lines, infrastructure) achieving ~81% mIoU on 40 large-scale urban scan tiles. Engineered a full preprocessing pipeline with Open3D and Rasterio - voxel downsampling, normal estimation, 3D patch extraction from raw .las files - and deployed inference as a FastAPI REST endpoint returning class-annotated 3D point clouds.

PyTorch PointNet++ Open3D CUDA Rasterio FastAPI

View on GitHub

Diabetic Retinopathy Severity Classification

Fine-tuned EfficientNet-B4 (pretrained on ImageNet) on the APTOS 2019 dataset for 5-class DR severity grading, implementing an OpenCV-based preprocessing pipeline (CLAHE contrast enhancement, spatial normalization, augmentation) with GPU-accelerated training achieving ~86% accuracy and a 0.87 Quadratic Weighted Kappa. Applied Grad-CAM visualization to generate lesion-level activation heatmaps, enabling clinically interpretable predictions.

PyTorch EfficientNet-B4 OpenCV CUDA Transfer Learning

View on GitHub

Exoplanet Classifier - NASA KOI, K2 & TESS

Built and deployed an end-to-end ML pipeline for automated exoplanet detection using supervised learning on NASA KOI, K2, and TESS datasets, exposed as a web application on AWS EC2.

Certifications

Achievements

NASA Space Apps Challenge - Global hackathon building solutions on satellite & geospatial data problems
Smart India Hackathon (SIH) - India's largest national-level government-backed hackathon, representing Symbiosis Institute of Technology
Symbiothon - Institute-level hackathon, Symbiosis Institute of Technology
Google Campus Ambassador - Represented Google on campus promoting developer programs
Senior Editor, Department Magazine - Led editorial planning and contributor coordination

EDUCATION

B.E. in Artificial Intelligence & Machine Learning

Symbiosis Institute of Technology

2024 - 2028

Pune, India

LANGUAGES

English

Hindi

Marathi

GET IN TOUCH

Let's build something amazing together

LinkedIn GitHub Email