Building AI that sees, understands, and interacts.
Scroll to explore
Ayush Shukla
AI/ML Engineer • Computer Vision • AR/VR
technospes / AURA
technospes / AgriFixAR
technospes / StrikeSync
technospes / BharatLok AR
technospes / MapARly
technospes / ParleyAlly
technospes / DistribuFlow
technospes / CoachSaab
technospes / DistribuFlow
technospes / ParleyAlly
technospes / MapARly
technospes / BharatLok AR
technospes / StrikeSync
technospes / AgriFixAR
technospes / AURA
technospes / CoachSaab
About Me
I am an AI/ML Engineer specializing in real-time Computer Vision (30-60 FPS), production-ready RAG systems (<100 ms retrieval), and highly reliable backend pipelines. I focus heavily on writing production-grade software that bridges the gap between hardware constraints and advanced machine learning models.
With deep exposure across computer vision frameworks, generative model tuning, and low-latency rendering, I design intelligent agents that seamlessly interact with real-world environments through touchless controls and scalable architectures.
EDUCATION
Galgotias College of Engineering and Technology
Bachelor of Technology in Computer Science & Engineering
Expected graduation: May 2027 | Uttar Pradesh, India
Relevant Coursework: Artificial Intelligence, Data Structures & Algorithms, OS, DBMS, Probability & Statistics, Linear Algebra.
TECHNICAL SKILLS
AI/ML: PyTorch, YOLO, OpenCV, MediaPipe, Whisper, Hugging Face, scikit-learn, NumPy, Pandas
Model Training & Evaluation: QLoRA, LoRA, SFT Data Prep, DPO Dataset Construction, Hugging Face Trainer, MLflow
Vector Databases & RAG: Pinecone, ChromaDB, FAISS
Backend & Systems: FastAPI, Multithreading, CUDA, WebSockets, UDP, orjson, Docker, Rate Limiting
Languages & Graphics: Python (Expert), C++, C#, C, Unity 3D
My Projects
02
AURA — Vision-Language AI Agent
• Built a real-time multimodal AI agent integrating vision (OpenCV + MediaPipe) and voice (Whisper + LLM APIs), achieving a smooth 60 FPS profile using threaded pipelines and non-blocking asynchronous inference systems.
• Implemented gesture-driven OS navigation (cursor control, typing, application management) configured with temporal smoothing and adaptive noise filtering to retain precision across rapid motion.
• Formulated an asynchronous voice and tool-calling infrastructure running concurrent execution states within strict standard laptop memory ceilings (<4 GB RAM, CPU fallback).
AgriFix AI — Multimodal Repair Assistant
• Developed a multimodal AI processing pipeline (Whisper → MobileCLIP → ChromaDB → Gemini) delivering dynamic, context-aware mechanical guidance generated directly from audio-visual farm inputs.
• Highly optimized vector retrieval layers (<100 ms) and reliable structured extraction schemes capable of filtering out extreme environmental noise and adversarial prompt-injections.
• Containerized backend services built on FastAPI using robust input-validation structures for portable, deterministic deployments via Docker.
CoachSaab — CV Based Form Corrector
• Engineered an AI-powered fitness and sports coaching mobile application using real-time computer vision, pose estimation, and biomechanics-based analysis to detect movement quality and provide instant form corrections.
• Implemented on-device pose tracking and deterministic form analysis for real-time exercise, yoga, and sports movement feedback.
• Built a Flutter + FastAPI architecture with PostgreSQL/Supabase and Groq-powered AI for personalized coaching, workout insights, and performance reports.
DistribuFlow — AI-Powered Order Management
• Developed agentic AI-powered distributor order management system automating conversational order processing through WhatsApp in Hindi, English, and Hinglish.
• Built LangGraph-based agent with deterministic product resolution, fuzzy matching, and confidence-based routing to prevent SKU hallucinations.
• Implemented complete order lifecycle with credit validation, manager approval workflow, and automated PDF invoice generation via n8n automation.
StrikeSync — Real-Time Motion Capture Engine
• Architected an ultra-low latency skeletal tracking motion pipeline streaming motion vectors directly across diverse engines (Python → Node.js → Unity) at 30+ FPS without proprietary hardware tracking.
• Leveraged CUDA-accelerated YOLO models paired with intelligent adaptive frame skipping, optimizing workflows for both discrete high-end GPUs and CPU edge fallbacks.
• Implemented lightweight serialization over high-frequency UDP pipelines integrated with custom threshold data-smoothing to neutralize frame drops and network jitter.
BharatLok AR — Cultural Heritage AR Experience
• Developed a mobile AR application transforming Indian heritage site visits into immersive experiences with photorealistic 3D reconstructions overlaid on real monuments.
• Integrated Google Gemini API for conversational AI tour guide, enabling natural language queries about historical sites.
• Implemented Google Cloud Anchors for shared multi-user AR experiences and Firebase for dynamic content delivery.
MapARly — Geospatial AR Navigation
• Engineered an immersive AR navigation system combining Mapbox geospatial data with on-device YOLOv5n object detection for contextual environmental awareness.
• Implemented GPU-accelerated inference via ONNX Runtime achieving real-time obstacle and object detection on mobile devices.
• Developed spline-based 3D path rendering with object pooling and frustum culling for optimal mobile performance.
ParleyAlly — Legal Document Demystifier
• Built a Python-based backend service using OCR and Generative AI to transform complex legal documents into plain-language summaries.
• Implemented RAG pipeline with PyPDF2 and Tesseract OCR for universal document ingestion and hallucination prevention.
• Designed scalable Flask REST API for seamless integration with frontend applications.
1 / 3
Experience
03HexSoftwares
AI Engineer Intern | Remote, India (Oct 2025 - Dec 2025)
• RAG Pipeline Architecture: Designed production RAG topologies (Pinecone + Hugging Face) yielding a 42% reduction in end-to-end query latency (120 ms → 70 ms) via batch document ingestion algorithms and parallelized retrieval logic.
• Fault-Isolated Systems: Engineered asynchronous backend services splitting text embedding, vector lookups, and model inference into decoupled modules guarded by robust fallback handlers and strict request constraints.
Honors & Achievements
• 1st Place Winner — HackOverflow v9.0 (2025) outperforming 600+ national competitors with an AI-powered AR system.
• National Finalist — Smart India Hackathon (SIH) 2025.
• Finalist — India Innovates (Asia's largest hackathon arena, scaling across 20,000+ technical teams).
• Selected Finalist — CosmoHack 1, DataVerse, and BITS Pilani TechSeva competitive tracks.
• Top 5 — Axis HackSprint (National Technical Tier).
CONTACT
04Let's Connect
I am regularly open to technical discussions regarding real-time vision setups, optimization across high-throughput model backends, or immersive interactive graphics pipelines. Reach out via email or any of my active networking portals:
📧 Email: ayushshukla6917@gmail.com
Let's coordinate on industry-level AI integration paradigms or production engineering frameworks.
Github
Explore full open-source setups for low-latency trackers, multithreaded pipelines, and robust system code layers.
