Adrianos Botsios

A transformer inference engine written in C++ from scratch, with a cache-blocked NEON SIMD multithreaded matmul, a KV cache, and INT8 quantization, parity-tested against PyTorch. Together those took decode throughput from 0.2 to 88.4 tokens/sec. On top of the engine I wrote a JSON Schema to finite-state-machine compiler that masks logits at each decode step, taking a 360M model from 0/100 to 100/100 schema-valid tool calls at 0.037% of decode time.

An AI-native travel agency that plans and books trips end to end. A 10-agent LangGraph pipeline fans out across flight, stay, and date combinations, prunes over-budget branches, and re-plans whenever a critic step fails. The system runs on TypeScript, Node.js, Express, Zod, React, and Python behind a streaming REST API, bounding provider fan-out with a per-provider concurrency limiter, request coalescing, and a five-minute TTL cache on provider results. I also fine-tuned a local 1.5B model (LoRA on MLX) to cut inference costs, evaluated on schema validity, rank agreement, and per-slice error.

Awarded a Brown Data Science Institute grant to research data labeling and curation for vision-language-action (VLA) models, leading a 7-member team across CS, linguistics, and psychology studying multimodal data for emotion-aware humanoid robots. I built a FAISS vector database over 1.2M embeddings and showed that label-error detectors fall to chance under correlated corruption, and produced the first inter-annotator agreement measurement on DROID's labels (37.6K episodes, α = 0.813).

Worked on applied machine learning problems across healthcare and recommendation systems, including a breast cancer detection model using dimensionality reduction and ensemble methods, and a movie recommendation system built on statistical modeling and matrix factorization.

Built multi-agent orchestration systems end to end in Python for internal tooling, ticketing, data analysis, and aggregation pipelines, integrating internal APIs, deploying containerized services, and helping shape AI policy alongside executives. I also worked on Upstream's Grow product, building full-stack features in TypeScript, React, and Next.js from Figma designs with senior engineers and designers, handling authentication and security with unit tests and CI/CD.

Worked at the intersection of AI and policy, contributing to discussions on balancing innovation with ethical and regulatory frameworks. Provided input toward the Convention on Artificial Intelligence, Human Rights, Democracy, and the Rule of Law, with a focus on responsible deployment and long-term societal impact.

Worked across both software and robotics systems, building a full-stack IoT platform for smart home automation and improving object detection models for industrial robotics, including integration into a live production environment.

Developed an NLP system to analyze Greek-language customer feedback and extract satisfaction signals, supporting internal business intelligence efforts, while contributing to product development and team coordination.

Building systems that allow businesses to automate operations through AI agents, from communication and data entry to full workflow execution. Focused on making automation practical, deployable, and aligned with real business processes.

Selected as one of 20 students for Innovation Dojo, a competitive startup accelerator, completing the full venture-building process from idea to pitch. Developed and pitched an AI-powered grading and feedback platform for educators that placed first in the final competition.

Created and led a multi-year initiative bringing AI education and ethics into public schools through curriculum development and hands-on AI labs. Reached over 12,000 students and collaborated with research institutions and international organizations.

Led a 23-member team focused on applying AI and robotics to real-world problems. Raised funding, built working systems, and developed both hardware and software across projects, emphasizing complete, end-to-end solutions.

Led a 21-member team centered on sustainability and technology, designing and engineering large-scale smart city models integrating AI, IoT, and renewable energy systems. Raised over €10,000, published work on the project, and collaborated with local municipalities to explore real-world implementation.