First fully natural language GTM tool - your AI GTM engineer for go-to-market strategy.

Built because GTM planning requires synthesizing complex market research and competitive analysis. Hermes combines multiple AI models to handle market research, competitor analysis, and strategic planning. The tool generates comprehensive GTM strategies from natural language product descriptions, conducting automated market research and creating actionable execution plans.

Downloaded and embedded 2M AI research papers from arXiv into Pinecone VectorDB with RAG for semantic paper search.

Academic research discovery is broken - researchers waste hours finding relevant papers and understanding connections between work. I built a system that processes 2 million arXiv papers, creates embeddings for semantic search, and uses RAG to provide intelligent summaries. The custom PDF viewer includes highlighting, annotation, and automatic citation generation. The system can answer complex queries like 'show me papers combining transformers with RL' and provide contextual summaries of the most relevant work.

RAG app solving deprecated API usage by scraping docs for popular ML libraries and providing real-time API reference.

Developers constantly struggle with deprecated APIs and outdated documentation. Tribunary scrapes and indexes documentation for 50+ popular ML libraries, tracks API changes over time, and provides intelligent migration suggestions. The system can detect when you're using deprecated functions and suggest modern alternatives with code examples. It's like having a constantly updated reference that understands the evolution of ML libraries and can guide you through breaking changes.

A high-performance Python library for batch processing and concurrency of LLM API calls, achieving up to 1000x faster speeds than sequential API calls.

Built for high-volume offline data processing with concurrency, dynamic batching, checkpointing, and caching that reduces multi-day jobs to minutes. Handles real-time request handling with dynamic concurrency, load balancing, and rate limiting to keep latencies low and handle spikes. Perfect for agent or multi-step reasoning workflows that need to manage repeated LLM calls in a single user flow or tool + LLM pipeline. Also great for synthetic data generation to quickly produce large amounts of synthetic text for training or testing.

Built a product that processes 29M YouTube videos in real-time, extracting sponsor relationships with 92% accuracy.

The influencer marketing ecosystem lacks transparency around sponsor relationships. I built a distributed system that processes 29 million YouTube videos, extracting sponsor mentions with 92% accuracy using custom LLM parsing. The system handles 100K tokens per minute through intelligent batching and concurrency optimization. The analytics dashboard tracks 3K+ brands and 7K creators, revealing hidden patterns in the $16B influencer market. We partnered with creators totaling 600k+ followers, but learned the real value was in the data insights rather than lead generation.

Built an end-to-end CNN + LSTM pipeline that learns motion patterns directly from camera data, outperforming ORB-SLAM by 40% in dynamic scenes.

Traditional SLAM methods like ORB-SLAM fail catastrophically in dynamic environments and low-texture scenes. I developed an end-to-end deep learning approach that learns visual odometry directly from camera data without hand-crafted features. The CNN extracts spatial features while the LSTM captures temporal motion patterns. After building custom evaluation testbeds, the system outperformed ORB-SLAM by 40% in dynamic scenes and 25% in featureless environments. The work won the Massachusetts State Science Fair Grand Prize ($10k) and led to extended research at CMU AirLab integrating IMU data through Kalman filtering.

Custom learning management system built for AI education with interactive coding environments and automated grading.

Existing LMS platforms weren't designed for hands-on AI education. I built a custom system with real-time code execution environments, automated grading for Python and ML assignments, and interactive curriculum delivery. The platform includes features like live coding sessions, automatic plagiarism detection, and personalized learning paths. The 95%+ completion rate across 150+ students proved that the right educational technology can dramatically improve learning outcomes.

Intensive program focused on autonomous systems where I led a team project on reinforcement learning for self-driving cars.

This competitive MIT program provided deep exposure to autonomous systems development. I learned control systems, computer vision, localization, and path planning from leading researchers. The capstone project involved developing reinforcement learning algorithms for autonomous vehicles navigating complex environments. The experience provided crucial foundations in robotics that directly influenced my later work on DeepVO and research at CMU AirLab. The program's emphasis on real-world applications taught me how to bridge the gap between academic research and practical robotics systems.

Worked under a Stanford PhD student to implement various ML models for stock prediction, focusing on LSTM + attention mechanisms.

After losing money day trading, I got serious about systematic approaches to markets. Working under a Stanford PhD student, I implemented LSTM + attention models for directional market prediction rather than price forecasting. We conducted extensive hyperparameter optimization and backtesting across different market regimes. The research taught me that most 'profitable' strategies don't survive real market conditions due to overfitting and regime changes. The work expanded into algorithmic sports betting and arbitrage strategies, providing valuable lessons about building robust systems in adversarial environments.

Built a marketplace connecting high school tutors with younger students during COVID, onboarding 300+ tutors.

COVID created massive demand for remote tutoring while existing platforms were expensive and impersonal. I built a two-sided marketplace from scratch that made it easy for high schoolers to offer tutoring and parents to find local help. The platform handled tutor verification, scheduling, payments, and quality control. We onboarded over 300 tutors and facilitated thousands of tutoring hours. This was my first experience building marketplace dynamics, user acquisition funnels, and payment processing systems. The project taught me the fundamentals of building trust in educational technology platforms.

Led programming team to create comprehensive autonomous control system, ranking 4th globally out of 7,000+ teams.

Competitive robotics demands systems that work reliably under extreme pressure. I led a 3-person programming team building a complete autonomy stack: computer vision for object detection, dead-wheel and visual odometry for localization, PID controllers for precise movement, and custom motion profiling for smooth trajectories. Our autonomous routines ranked 4th globally out of 7,000+ international teams. We achieved a 36-1 season record with state championships in both Massachusetts and New Jersey, plus 'Best Autonomous' and 9 other awards. The experience taught me how to build robust systems that perform when it matters most.

Voice-first Android app for visually-impaired users with custom CNN encoder + RNN decoder for image captioning and scene description.

Existing accessibility tools were clunky and didn't leverage modern AI advances. I built a comprehensive voice-first Android app addressing navigation and recognition challenges for visually-impaired users. The system features a custom CNN encoder + RNN decoder for image captioning, scene description, emotion recognition, and robust OCR - all accessible through intuitive voice commands. I worked directly with visually-impaired users throughout development to ensure practical utility. The app attracted attention from 30+ investors at Boston University's BUild Lab and taught me how to build AI that works for real users in real-world conditions.