Akhil Nelaturi

Built a regression model based on a convolutional neural network (CNN) for a self-driving car. This autonomous driving model predicts the steering angle to be turned when viewing an image of the road.

Led the development of a segmentation pipeline for aerial geo-spatial imagery to automate landscape measurement tasks. The project focused on reducing manual annotation effort and improving scalability for large-scale mapping and analysis.

I utilized HR-Net, a high-resolution convolutional network, to achieve precise segmentation of landscape features such as vegetation, built structures, and terrain boundaries. The model was trained on custom-labeled datasets and optimized for generalization across varied geographic regions.

Key Contributions:
• Reduced manual annotation effort by 70% through high-fidelity segmentation outputs.
• Deployed the model using AWS ECS with Fargate, enabling scalable, containerized inference across multiple regions.
• Integrated preprocessing pipelines for image normalization, tiling, and geo-referencing.
• Designed evaluation metrics for segmentation accuracy and spatial consistency.
• Collaborated with GIS analysts to validate outputs and align model predictions with real-world measurements.
• Documented the full ML lifecycle, including training scripts, deployment configurations, and performance benchmarks.

Designed and deployed an end-to-end system for detecting traffic signs and tagging their geo-locations using satellite and street-level imagery. The solution leveraged a custom-trained Faster R-CNN (FRCNN) object detection model, built on a dataset curated specifically for varied traffic sign types and environmental conditions.

The model was hosted on an AWS EC2 instance and integrated with geospatial tagging logic to associate detected signs with precise coordinates automatically. This automation reduced manual annotation efforts by over 80%, significantly accelerating infrastructure mapping and compliance workflows.

Key contributions included:
• Data preparation and annotation for diverse traffic sign categories.
• Model training and hyperparameter tuning for robust detection across lighting and occlusion scenarios.
• Geo-location tagging using image metadata and external mapping APIs.
• Scalable deployment on cloud infrastructure for batch and real-time processing.

This project demonstrates my ability to combine computer vision, geospatial data handling, and cloud-native deployment to solve real-world automation challenges.

Built a comprehensive system to automate the extraction of structured information from both typewritten and handwritten forms. The pipeline included object detection to locate relevant fields in typewritten documents and bounding boxes for handwritten characters. A classification model was trained on top of these detections to recognize individual handwritten characters.

I used open-source OCR tools for typewritten text and developed a custom CNN model trained on E-MNIST to classify handwritten inputs. To enhance training data quality, I created a PyQt-based UI for manual labeling of handwritten characters, enabling real-time annotation and dataset expansion.

This project demonstrates expertise in document processing, OCR, computer vision, and UI-driven data labeling workflows.

Led the development of an automated alt-text generation pipeline tailored for scientific publishing workflows. The system leveraged OpenAI and Gemini Flash models to generate descriptive text for diverse image types, including charts, graphs, biological illustrations, chemical structures, and mathematical equations.

I built a custom image classification model to identify asset types and route them through subject-specific few-shot prompts, improving contextual relevance and semantic accuracy. I also integrated a feedback-driven regeneration loop using chatbot-style interaction, allowing editors to refine outputs interactively and enhance accessibility compliance.

This project significantly improved editorial efficiency and accessibility standards across scientific content.

Developed a domain-specific GenAI chatbot to assist research teams in comparing and analyzing drug trial data. The system was built using a Retrieval-Augmented Generation (RAG) architecture, with automated ingestion and retrieval powered by AWS Lambda and Amazon Kendra, enabling scalable and context-aware responses.

I integrated multimodal models to interpret visual elements in clinical documents—such as charts, dosage graphs, and trial timelines—enhancing the chatbot’s ability to respond accurately across diverse data formats. I curated prompt templates and retrieval filters to align with medical terminology and regulatory standards.

I also evaluated retrieval quality using Precision@k, and benchmarked LLM responses with ROUGE and BLEU scores against SME-written answers, ensuring scientific rigor and reliability.

This project demonstrates expertise in GenAI, multimodal integration, and pharmaceutical domain adaptation for enterprise-grade chatbot systems.

Led the development of a GenAI-powered automation system for call center workflows, focused on transcription, compliance, and intelligent agent assistance. The pipeline began with audio transcription using OpenAI’s Whisper, enabling downstream NLP tasks and searchable conversation logs.

I integrated GLiNER, a transformer-based generalist NER model, to detect and redact personally identifiable information (PII) from transcripts, ensuring regulatory compliance (e.g., GDPR, HIPAA). I built a Retrieval-Augmented Generation (RAG) system using open-source LLMs and product-specific embeddings to automate agent responses for a B2B platform, reducing average handling time and improving customer satisfaction.

The system was designed for real-time use, with scalable deployment and modular components for easy adaptation across industries.