Ashish Kushwaha

Advised startups and enterprise teams on Voice AI and GenAI adoption, evaluating business use cases, technical feasibility, technology stacks, and scalable AI product roadmaps. Conducted architecture reviews for real-time conversational AI systems, guiding teams on WebRTC, streaming STT/TTS, low-latency LLM pipelines, RAG frameworks, and cloud deployment strategies. Designed Voice AI solution architectures involving speech-to-text, LLM reasoning, retrieval-augmented generation, text-to-speech, telephony integrations, and human handoff workflows. Guided teams on LLM inference optimization using hosted open-source models, vLLM serving, KV cache reuse, request batching, token streaming, and latency-aware deployment patterns. Reviewed GenAI system designs for agent orchestration, tool calling, memory management, retrieval quality, hallucination reduction, observability, and production readiness. Mentored engineering teams on best practices in LLM integration, Voice AI pipelines, agentic workflows, cloud deployment, and scalable system design for production-grade AI applications.

Provided technical consulting on enterprise-wide AI adoption, accelerating AI-driven transformation strategies. Designed and deployed a full-stack RAG pipeline on Azure, enhancing retrieval accuracy and response efficiency. Eliminated manual data searching by implementing a RAG pipeline that delivers precise, instant answers to sales teams saving hundreds of hours, enhancing customer interactions and decision-making.

Led a team of 4 engineers to deliver a low-latency RAG chatbot for accurate cyber security knowledge retrieval and threat intelligence support. Built a full-stack RAG LLM solution for a cyber security platform, integrating document ingestion, vector retrieval, LLM generation, citations, and user-specific chat sessions. Fine-tuned LLaMA-3.1 for cyber security and threat intelligence use cases, improving domain-specific response quality for malware, vulnerabilities, alerts, and incident context. Implemented RAG pipelines with fine-tuned models to retrieve accurate and precise real-time threat intelligence data from cyber security knowledge sources. Hosted open-source LLMs using vLLM inference server, improving serving throughput with continuous batching, optimized KV cache reuse, and efficient GPU memory utilization. Optimized LLM inference latency using KV caching, PagedAttention, request batching, token streaming, and model serving configuration for production RAG workloads. Built stateful agent orchestration using LangGraph, enabling conditional reasoning, contextual memory, tool routing, and user-specific responses in production RAG systems. Implemented Redis cache and memory management to store short-term conversation memory, session state, and reusable LLM query-response context. Implemented session management using Redis to isolate user-specific requests, chat history, and retrieval context across concurrent users. Built CI/CD pipelines, deployed services on AWS EC2, and implemented LLMOps practices including model serving, monitoring, versioning, and production deployment workflows.

Modeled RAG workflows as a state graph using LangGraph, enabling deterministic control over multi-step LLM reasoning and tool execution. Implemented Query Transformation Feature to enhance input user query to get accurate retrieved documents. Implemented In-text Citation Documents feature using Prompt Engineering. Implemented Claude3 Haiku model using AWS Bedrock. Developed Data Ingestion Pipeline between AWS S3 Knowledge Base, AWS Kendra and AWS ECS.

• Developing a full RAG architecture to deliver a working AI model that can answer their usecase related queries. • Implemented pipelines to retrieve the documents from AWS Kendra and proceed to pass it to LLM model. • Developed Data Ingestion Pipeline between AWS S3 Knowledge Base, AWS Kendra and AWS ECS. • Implemented a data cleaning script to split large pdfs into smaller chunks and clean the raw data. • Implemented Query Transformation Feature to enhance input user query to get accurate retrieved documents. • Thoroughly studied and experimented with multiple models to choose the suitable one for our use case. • Improving model's performance by checking their resultant matrix using tools like @uptrain.ai. • Implemented In-text Citation Documents feature using Prompt Engineering.

Voice bot Developer for HealthCare service: Built a chatbot and real-time voicebot using LiveKit + WebRTC to handle patient queries and healthcare support. Fine-tuned advanced NLP models to improve the systems understanding and accuracy for medical-related queries. Developed a symptom-checker module to help users identify potential health issues based on their described symptoms. Implemented a low-latency STT LLM TTS pipeline, enabling smooth and natural voice interactions for the voicebot. Optimized concurrency and performance, ensuring reliable and scalable operation during peak patient interaction hours.

Cancer Cells Detection: Gathered a comprehensive dataset of medical images, including histopathological slides, mammograms, or radiological scans. These images are annotated with information about the presence or absence of cancer cells. Chose a suitable deep learning architecture for image classification and detection tasks, such as Convolutional Neural Networks (CNNs). The models ability to recognize subtle patterns and anomalies is crucial. Conducted rigorous validation and testing to evaluate the models performance. Calculated essential metrics such as accuracy, sensitivity, specificity, and the area under the Receiver Operating Characteristic (ROC) curve. Deployed the trained deep learning model for practical application, integrated into a healthcare system. This enables real-time or batch analysis of medical images for cancer cell detection.

Waste in the Street Photo Detection: Developed algorithms that could effectively identify and quantify waste in various environments captured in photographs. Utilized a custom CNN model for image classification, capable of identifying and categorizing waste items in street photographs. Created an automated system that processed images and flagged those containing street waste, enabling faster and more efficient waste detection. Potholes in Photo Detection: Developed a computer vision model to detect potholes in photographs, thereby aiding municipal bodies in road maintenance tasks. Created an automated system that streamlines the detection process, reducing the need for manual inspection and speeding up response times Improved response times for pothole repair by providing an automated system for pothole identification and prioritization Hyperboles in Radargram Detection: Designed a system that effectively detected hyperbolic patterns, in radargram images Demonstrated proficiency in radar data analysis, including data preprocessing, signal identification, and real-time processing. Trained the deep learning model using a well-annotated dataset, ensuring the model learned to recognize relevant patterns.