AI SYSTEMS & SOFTWARE ARCHITECT

Engineering complete AI systems from notebook to product.

I build production-oriented AI architectures — pairing computer vision models, geospatial machine learning, and high-throughput backend APIs into reliable software people can actually use.

Explore Builds Get in Touch

pratham@systems-core:~

$ cat profile.json

{
  "name": "Pratham Kashyap",
  "role": "AI Systems Engineer",
  "focus": ["Applied AI", "Computer Vision", "Backend APIs"],
  "flagship": "AthenaeumAI",
  "pipeline": "idea → model → api → data → system"
}

$ system-status --all

✓ 5+ Shipped Builds ✓ 4 Research Tracks ✓ Real-Time CV Deployed

Philosophy

Bridging the Gap Between Models & Production.

Training an algorithm in a Jupyter notebook is usually just the first 20%. The remaining 80% is where engineering matters — data validation, low-latency API contracts, resilient storage schemas, and fault-tolerant user interfaces.

End-to-End

Applied AI Systems

Building complete pipelines: chunk-based document ingestion, semantic vector retrieval, and LLM inference generation with strict output contracts.

Real-Time

Computer Vision

Lightweight edge-ready vision pipelines using YOLOv8/11 and SAM for object detection, dwell-time estimation, and crop pathology diagnosis.

Resilient

Backend & Distributed APIs

Designing RESTful services in Node.js, Express, FastAPI, and Java Spring Boot with clean schema validation and persistent databases.

Selected Work

Featured Engineering Builds.

A showcase of systems where machine learning models are wrapped into robust, production-ready software architectures.

BUILD // 01 · ACTIVE FLAGSHIP

AthenaeumAI

AI-powered study platform that transforms dense course material into interactive quizzes, active recall sessions, and adaptive revision schedules.

Engineered document parsing with chunk-based semantic extraction, Groq LLM-powered test generation, and MongoDB state persistence behind a React web interface.

React Node.js MongoDB Groq API Tailwind

View Source Code

ARCHITECTURE Semantic RAG Pipeline

GENERATION SPEED < 1.2s Latency

DATA LAYER MongoDB Atlas

FRONTEND Modular React

BUILD // 02 · GEOSPATIAL ML

CiviQ — Civic Issue Register

Geospatial machine learning platform for automated classification and severity prioritization of urban municipal complaints.

Combines domain-knowledge labels, KNN spatial density features, DBSCAN hotspot clustering, and an XGBoost ensemble classifier validated against real NYC 311 incident records.

Python XGBoost DBSCAN React Firebase

View Source Code

VALIDATION ACCURACY 95.7% Verified

DATASET SOURCE NYC 311 Open Data

CLUSTERING DBSCAN Hotspots

CLASSIFIER Tuned XGBoost

BUILD // 03 · COMPUTER VISION

Smart Retail Store Assistant

Zero-shot single-camera shelf monitoring, inventory dwell-time estimation, and real-time stock removal anomaly detection.

Built using YOLOv8 object detection paired with velocity-predicted centroid tracking to monitor specific shelf coordinate polygons without requiring custom domain re-training.

Python OpenCV YOLOv8 NumPy SciPy

View Source Code

CAMERA SETUP Single Monocular

DETECTOR YOLOv8 Real-Time

TRACKING METHOD Velocity Centroid

ALERT LATENCY < 80ms Response

BUILD // 04 · MULTIMODAL AI

Multimodal AI Assistant

Face-gated voice assistant that co-activates voice recognition only when user visual presence and attention are verified.

Integrates OpenCV Haar-cascade presence gating, OpenAI Whisper voice transcription, and a deterministic 5-category intent classification engine with fallback handling.

Python Whisper OpenCV MediaPipe Tkinter

View Source Code

GATING MECHANISM Visual Gaze + Face

AUDIO INFERENCE Whisper Base

FALSE POSITIVES Near-Zero Trigger

PUBLICATION IEEE Research Paper

BUILD // 05 · COMPUTER VISION

AgriTech — Crop Disease Detector

Computer vision pipeline combining the Segment Anything Model (SAM) and YOLOv11 for multi-crop leaf pathology segmentation and diagnosis.

Isolates irregular leaf surfaces using prompt-driven SAM segmentation masks, followed by custom-trained YOLOv11 and TFLite lightweight models for wheat, tomato, and cotton diseases.

YOLOv11 Segment Anything (SAM) TensorFlow Lite OpenCV

View Source Code

SEGMENTATION Meta SAM Masks

DETECTOR YOLOv11 Pathology

SUPPORTED CROPS Wheat, Tomato, Cotton

EDGE DEPLOYMENT TensorFlow Lite

Tech Stack

Interactive Technology Constellations.

Core engineering domains mapped as dynamic celestial geometries. Hover over any star to illuminate its orbital conduits.

01 // Languages

Python C++ Java TypeScript JavaScript SQL

02 // AI & Vision

PyTorch OpenCV YOLOv8/11 SAM Scikit-learn Whisper

03 // Backend & Web

Node.js FastAPI Express Spring Boot React Vite

04 // Data & Cloud

MongoDB PostgreSQL Docker Git Firebase Linux

Research Notes

Applied ML & Empirical Signals.

Structured research findings in neuro-fuzzy ranking systems, geospatial municipal triage, and acoustic biomarker detection.

TELEMETRY // 01 R² = 0.98 FUZZY DECISION MODEL

GA-ET-IVCFS-ANFIS Ranking Architecture

Interval-valued complex fuzzy adaptive neuro-fuzzy inference system (ANFIS) optimized with genetic algorithms for robust multi-criteria decision-making under uncertainty and noise.

Complex Fuzzy Logic Genetic Optimization ANFIS Systems

TELEMETRY // 02 95.7% ACC GEOSPATIAL TRIAGE

Civic Complaint Prioritization Framework

Hybrid synthetic + NYC 311 evaluation framework combining KNN density clustering and tree ensembles for automated geospatial municipal incident classification.

XGBoost Ensemble DBSCAN Spatial Hotspots NYC 311 Verified

TELEMETRY // 03 F1 = 0.935 ROC-AUC = 0.945

Parkinson's Acoustic Biomarker Mining

Acoustic feature extraction and high-dimensional SVM classification pipeline for non-invasive voice disorder detection and early screening signals.

Acoustic Signal Extraction SVM Kernel Classifier Biomarker Triage

TELEMETRY // 04 IEEE SOLO PUBLISHED PAPER

Multimodal Gaze & Voice Interaction Systems

Research examining low-overhead presence-gated audio co-activation architectures for responsive, false-trigger-resistant human-computer multimodal interfaces.

Haar Face-Gating Whisper Speech Base Zero False Triggers

Space Arcade

Asteroid Defender.

Interactive canvas mini-game built right into the portfolio. Dodge deep space asteroids and gather energy stars!

SCORE: 0 BEST: 0

Mission Briefing

Pilot your craft with arrow keys or WASD. Dodge rogue asteroids and collect cyan stars.

Controls: ▲ ◀ ▼ ▶ or WASD

Let's build software that matters.

Whether you want to discuss AI systems, computer vision pipelines, backend architecture, or potential collaborations — my inbox is always open.

Email Me Directly

✓ Email copied to clipboard!