Pranav Unnikrishnan

ml · data · computer vision · generative ai
projects

End-to-end ML systems across domains.

Data science, data engineering, computer vision, robotics, and GenAI, with clean repos and measurable outcomes.

Selected projects

Filter by domain, or browse everything.

11 projects

Autonomous RC Car

Robotics SLAM • Planning GitHub ↗
  • Built an autonomous hallway-navigation system using LiDAR-based SLAM, RF2O odometry, and a ROS2 + ArduPilot pipeline.
  • Implemented a PD-based waypoint follower with dual-criteria waypoint advancement for reliable multi-lap navigation and accurate cornering.
  • Developed a LiDAR-only wall-following controller using RANSAC line extraction + geometric distance/orientation estimation for map-free navigation.
  • Created an A* path-planning pipeline: map preprocessing, padding, distance-field cost shaping, and waypoint discretization.
  • Prototyped a NumPy-only camera navigation system and evaluated a custom ICP-EKF localization approach.

Social Navigation Project (SocNavGym)

Robotics Reinforcement Learning GitHub ↗
  • Extended SocNavGym with generalized environment handling (padding + attention-based encoders) and a proxemics-based reward function.
  • Trained DQN agents across simple and complex social environments and compared reward models (Baseline, DSRNN, TGRF, Proxemics).
  • Ran large-scale evaluation (100-episode rollouts per agent) plus human A/B preference tests, where proxemics-based policies performed best in complex scenes.

Algorithms for Exploration and Exploitation in Reinforcement Learning

Reinforcement Learning Research / Implementations GitHub ↗
  • Implemented and compared value-based (Q-learning, SARSA, DQN) and policy-based (REINFORCE, PPO, DDPG) algorithms.
  • Designed and trained agents in classic control + gym environments (Cliff Walking, Taxi, LunarLander-v2, Pendulum).
  • Applied TD learning, Bellman equations, policy gradients, and actor-critic methods to study exploration-exploitation trade-offs.
  • Produced training curves and performance evaluations showing convergence and stability across algorithms.

Sawyer Robot Analysis

Robotics MuJoCo • RL GitHub ↗
  • Analyzed Sawyer task data with Python, OpenCV, and NumPy, and visualized kinematics, trajectories, and control signals for benchmarking.
  • Implemented Q-Learning, PPO, and DDPG in Sawyer (MuJoCo) environments for robotic grasping and trajectory-following tasks.
  • Evaluated convergence and stability to compare training behavior across algorithms.

DRL-Based Map-less Crowd Navigation for Mobile Robots

Robotics TD3 • SAC Gazebo • TurtleBot3
  • Designed and implemented a deep RL navigation system for autonomous mobile robots in crowded environments.
  • Developed reward functions incorporating collision probability (CP) and trained models with TD3 and SAC using TurtleBot3 + 2D laser scans in Gazebo.
  • Achieved strong gains: TD3 + CP reached 1387 successful episodes out of 1400 vs 255 without CP.

Real-Time AI-Enabled Retail Analytics & Data Pipeline

Data Engineering Streaming • CV GitHub ↗
  • Architected an edge-to-cloud pipeline (Jetson Nano → Kafka → PySpark → MongoDB Atlas) with real-time object detection for automated checkout and invoice generation.
  • Built an ALS-based recommendation engine and an MQTT bridge for personalized suggestions and high-throughput async transfer.

Football Player Performance Scraper & Analysis

Data FastAPI • Pandas GitHub ↗
  • Scraped and engineered a 64-feature dataset of 12,900+ players across 5 seasons (2020 to 2025) using Python, FastAPI, Pandas, and NumPy.
  • Applied EDA, trend modeling, and visualization to predict market value fluctuations and identify top performers for transfer analytics.

ST-RAG (Summarization Tree Retrieval Augmented Generation)

GenAI RAG
  • Developed a novel pipeline to index and query large document corpora by building a hierarchical tree structure through recursive text clustering and summarization. This enabled efficient information retrieval at multiple levels of abstraction, mitigating context loss in traditional, isolated text chunks.
  • Designed an advanced retrieval backend using UMAP to reduce high-dimensional text embedding's dimensionality. Integrated GMMs for soft clustering of text segments, dynamically determining the optimal cluster count using BIC.
  • Benchmarked system performance using the RAGAS framework to evaluate component-level metrics. Achieved an Answer Relevance score of 0.978 and a Context Recall score of 0.980, outperforming traditional BM25 and DPR baseline methods across diverse large language models.

Hyperpartisan News Classification (BERT + Longformer)

NLP Transformers GitHub ↗
  • Compared chunk analysis, summarization, and stride chunking to detect extreme bias in 10,000 news articles using BERT and Longformer.
  • Reached 95.5% accuracy using RoBERTa + chunk analysis, outperforming full-sequence Longformer.
  • Analyzed misclassifications and label frequency distributions to improve generalization.

Efficient Fine-Tuning of LLMs + RLHF

GenAI QLoRA • PEFT GitHub ↗
  • Fine-tuned Llama 7B on Vicuna (20,000 samples) using QLoRA and PEFT for parameter-efficient adaptation.
  • Explored optimization strategies including 8-bit matmul, NF4 quantization, and nested quantization.
  • Integrated RLHF via TRL + PPO and deployed a text-generation policy model behind a functional UI.

NBA Player Scoring Analytics & Regression Modeling

Analytics R • Regression GitHub ↗
  • Cleaned and processed NBA scoring + salary datasets in R to prepare analysis-ready data.
  • Performed EDA to identify trends, outliers, and differences between normal and high scorers.
  • Built and compared linear regression, Poisson regression, and GAM models to study scoring patterns.