VamanRajagopal

Camera-guided robot navigation with a streaming VLM and Arduino motor control.

• Integrated the Overshoot streaming VLM API with a live camera feed to translate scene observations into discrete motor commands for centering, approaching, stopping, and continuing through navigation targets.
• Built a browser-to-Arduino control layer using WebSockets and serial commands, with manual override, automatic reconnection, and per-inference latency logging to support safer debugging and repeatable hardware testing.

Visual node-graph builder for LLM conversational workflows.

• Implemented configurable nodes for OpenAI calls, external API requests, and information extraction from text and images, allowing nontechnical users to assemble agent-like workflows without writing orchestration code.
• Implemented node-level message isolation in a React, TypeScript, and Python conversational framework, restricting each node to relevant user context to prevent cross-flow contamination and reduce hallucination risk.

Classical, neural, embedding, and LLM classifiers compared on student responses.

• Built and benchmarked logistic-regression, neural-network, embedding-based, and LLM pipelines under a shared multiclass evaluation setup spanning classical, neural, semantic, and generative approaches on identical data splits.
• Achieved 80% test accuracy with an LLM classifier using structured function calling to constrain predictions to valid category labels.

Real-time game statistics platform and Discord bots.

• Built a real-time gaming statistics and community platform serving 60,000+ users, giving players access to in-game information without requiring them to remain active inside the game during live play sessions.
• Engineered FastAPI and Go backend services with MongoDB and SQL for a game serving 20,000+ concurrent players, handling player-stat queries, command traffic, and leaderboard updates during peak activity.

AR heads-up display for DJs on Snap Spectacles.

• Built an augmented-reality heads-up display for Snap Spectacles that presented live crowd-feedback signals to DJs so performers could monitor audience engagement without leaving the live mixing workflow.
• Developed a Flask API that streamed audio metadata from Rekordbox and a Pioneer DDJ-FLX4 controller into the Spectacles application, synchronizing track context with live crowd analysis in the wearable display.

Educational mini-games driven by model outputs.

• Implemented event-driven scoring and state-machine game loops for an educational mini-game system built during a hackathon, keeping gameplay behavior consistent across multiple activities while supporting different scoring and progression rules.
• Exposed machine-learning performance metrics through FastAPI endpoints consumed by the React frontend so model outputs could influence scoring, difficulty, and player-facing guidance throughout each learning session.

Volunteer coordination app for a nonprofit.

• Developed a volunteer-coordination application for Cube360 that centralizes upcoming events and RSVP information for nonprofit staff, volunteers, and community members across recurring service programs.