👋

Hello, I'm Naveen Prashanna. I'm an AI / Machine Learning Engineer with 3+ years of experience in building production AI/ML systems across quantitative finance, generative AI & browser automation.

Resume GitHub LinkedIn

My experience

Quantitative Researcher

Pharvision Advisers | USA

Engineered multi-factor machine learning models processing over 15 million daily tick-level data points across global equities and macro datasets, utilizing point-in-time alignment to achieve Sharpe > 1.2 at low portfolio turnover.
Orchestrated LLM-based multi-agent workflows using LangChain, MLflow, and W&B handling over 10,000 automated backtesting runs weekly across 50+ concurrent experiments, reducing model validation cycles from days to hours.
Developed an agentic production RAG system serving live financial research, combining Neo4j GraphRAG (multi-hop traversal) and dense vector search (FAISS) with cross-encoder reranking to process complex queries across 500GB+ of documentation.
Integrated a RAG evaluation framework utilizing LLM as a judge to benchmark response faithfulness, context relevance & retrieval accuracy.
Established scalable data pipelines with Polars, Airflow, and Delta Lake ingesting 10+ vendor feeds concurrently (handling 50M+ rows daily) to expand quantitative research coverage without latency degradation.

Jan 2026 – Present

AI Engineer

Kahana Group Inc | USA

Programmed production-ready LLM execution flows using Azure AI Foundry and Azure AI Services for an AI-native browser, orchestrating robust state transitions and prompt workflows to elevate reliability across complex multi-step user tasks.
Architected an Azure-backed persistence layer using Azure Blob Storage scaling to manage browser state, sessions, bookmarks, and tabs with sub-50ms sync overhead under heavy state transitions.
Instrumented real-time AI observability telemetry with Azure Monitor Application Insights, tracking token metering and usage analytics across stress tests simulating 5M+ monthly LLM throughput and trim cloud cost overhead.
Established automated evaluation harnesses using the Azure AI Evaluation SDK to analyze user interactions, model responses, and failure patterns at scale, boosting prompt optimization effectiveness and completion metrics.

Jul 2025 – Jan 2026

Software Engineer (ML Systems)

Quantitative Brokers | India

Built low-latency C++/Python inference runtimes and event-driven messaging pipelines executing predictive models with sub-millisecond dispatch (< 800µs) over high-throughput FIX protocol sessions processing millions of daily messages.
Integrated real-time feature streaming and state tracking for multi-leg execution strategies, dynamically processing live Level-2 order book feeds across 40+ asset pairs to minimize market execution slippage under peak institutional loads.
Engineered high-throughput PostgreSQL and Kafka telemetry pipelines capturing over 10GB of streaming fills and model predictions daily, eliminating data leakage and accelerating offline feature engineering by 3.5x.
Extended the core FIX messaging platform and backend services with multi-leg trade support, optimizing low-level data-access layers to handle heavy-load financial traffic spikes seamlessly.

Jul 2022 – Feb 2023

Data Scientist Intern

Big Data Science Research | Bangalore, India

Developed and deployed a machine learning pipeline using Python and scikit-learn to predict urban traffic patterns, enhancing model accuracy by 15% through the integration of geospatial data features.
Implemented a novel map-matching algorithm to refine GPS data, significantly improving the precision of traffic flow models by 20%.

Apr 2021 – Jun 2022

Machine Learning Engineer - Intern

Alphabt & TVS Motors Ltd

Implemented a program to scan & verify vehicle labels using OpenCV, boosting validation performance system by 3%.
Devised a custom TensorFlow-based object detection model, achieving 99% accuracy in text engraving recognition.

May 2019 – July 2019

My projects

Personal Assistant Bot

Made with:

Fine-tuned the Mistral LLM with QLora on a custom dataset to build a specialized and efficient personal assistant bot. Integrated a Retrieval-Augmented Generation (RAG) pipeline to query personal documents, ensuring precise answers.

GitHub

LLM-Powered Mac Automation Tool

Made with:

Developed a LangChain system for LLM Mac control, replicating core functionalities of Anthropic's Computer Control. Implemented tools for text simulation, mouse automation, image analysis, application management & web navigation.

GitHub

AI-Powered SQL Injection Detection System

Made with:

Developed a secure PHP-based web system with login, access control, and DistilBERT-powered SQL injection prevention. Achieved 99% detection accuracy using supervised fine-tuned LLMs; performed log analysis and penetration testing with SQLMap. Hardened the system using input sanitization, HTTPS setup, Apache WAF simulation, and prepared statements for robust security.

GitHub

BrainScanNet: MRI-Based Brain Tumor Detection with CNNs

Made with:

Built a CNN-based classification pipeline using VGG-16 and DenseNet to detect brain tumors from cropped MRI images. Preprocessed data via cropping and augmentation; VGG-16 model achieved highest accuracy among tested architectures. Designed an assistive tool aimed at enhancing radiologist efficiency and diagnostic consistency in tumor identification.

GitHub

Trajectory-Aware Human Feedback for Hierarchical RL

Made with:

Proposed a novel Hierarchical Reinforcement Learning framework to improve subgoal generation in complex tasks. Deployed the Deep-RL framework in the FetchReach environment, resulting in a 10% increase in task success rates

GitHub

Chatbot

Made with:

Built comprehensive knowledge base by web scraping & advanced NLP techniques enabling efficient data retrieval. Engineered an LSTM-based model with an attention mechanism, improving response relevance and context by 25%

GitHub

Surface Texture Analysis

Made with:

Systemized an approach in identifying machined surface textures with CV & ML techniques, achieved 99.6% accuracy. Built Neural Network employing statistical features from GLCM for texture classification improving accuracy by 44%

GitHub

Analysis of Optimization Algorithms

Made with:

Assessed the performance of various Stochastic Optimization algorithms on control agents created using OpenAI gym environment. Developed and applied multiple Gradient Descent algorithms to optimize Deep Q-Network which achieved a 15% boost in agent’s return

GitHub

Multi Agent Cricket Game - Reinforcement Learning

Made with:

Formulated cricket game as Markov Decision Process for optimal decision-making of actions in dynamic system. Modeled a 2-player Monte Carlo Tree Search algorithm for better action selection which elevated match outcomes

GitHub

Hangman Game

Made with:

Designed an algorithm for Hangman game where player has to guess letters of a word with limited number of guesses. Augmented N-gram language model for capturing letter patterns to improve prediction & achieved an accuracy of 62%

GitHub

Bike Tour Recommendation

Made with:

Optimized bike tour recommendations through exhaustive data preprocessing, data binning and feature engineering. Drafted an ensemble of XGB, LGBM and CatBoost with meticulous hyper-parameter tuning yielding a 0.71 accuracy

GitHub

Loan Default Prediction

Made with:

Developed loan default classifier using SMOTE and KNN Imputer for data pre-processing, achieving robust predictive accuracy. Achieved F1 score of 0.95 by rigorously tuning hyperparameters in Random Forest Classifier, significantly enhancing learning outcomes

GitHub

DIC’s Terrace Farming Robot

Made with:

Won Silver prize in DIC's Terrace Farming Robot for hilly areas challenge in Inter IIT Tech Meet among over 20+ teams. Designed an Autonomous agricultural robot that is capable of ploughing, seeding and harvesting by climbing up and down steps in hills. Programmed the bot with Arduino to follow the wall by a PID control system (Proportional-Integral-Derivative) using ultrasonic sensors

Unmanned Ground Vehicle

Made with:

Developed a program for UGV bot using ROS that transmits sensor & video feed to a remote system and tracked its path in the given map. Programmed the bot to create a 3D map of surrounding using kinect-360 and localize itself in the given map by Monte Carlo Localization

GitHub

Water Levitation Project

Made with:

Developed a program to control frequency of strobe lights and water pumps for levitating the water using the principle of Stroboscopic effect. Designed a PCB (Printed Circuit Board) using Eagle and fabricated the circuit along with Arduino to control pumps & lights to make patterns