Skip to content

About

GitHub profile repository

Resources

Readme
Activity

Stars

0 stars

Watchers

1 watching

Forks

0 forks
Report repository
BranchesTags

Latest commit

 

History

124 Commits

Folders and files

NameName
Last commit message
Last commit date
.github/workflows
.github/workflows
 
 
README.md
README.md
 
 

Repository files navigation

Csaba Toth - senior software engineer, remote team lead, public speaker, tech meetup enthusiast

Twitter Badge Linkedin Badge Instagram Badge GitLab Badge GitHub Badge

  • I use daily: JavaScript Python C# Flutter Git Shell VS Code Android Studio PyCharm Visual Studio
  • I work using: Django ASP.NET MVC Amazon AWS MySQL Azure GitHub GitLab Bootstrap GCP Linode Slack
  • I also use: Firebase Firestore Dart Java Kotlin HTML5 CSS3
  • I am also proficient: BigQuery PostgreSQL C++
  • Learning about: Machine Learning Docker Kubernetes

Generative AI and other Hackathon projects and wins

  • World Cup 2026 Live Predictor: Developed for the Elastic and AWS Hack Night (World Cup Edition), this project extends static tournament predictions with real-time in-game outcome forecasting. By capturing live match statistics from phone/TV screens during live gameplay and ingesting them into an Elasticsearch Serverless cluster, an agent performs momentum-based outcome predictions rather than relying solely on pre-match historical form.
    • Multi-index architecture on Elasticsearch Serverless: wc2026_matches storing tournament fixtures and results alongside wc2026_live_stats tracking per-minute in-game telemetry snapshots (possession, shots on target, passes, tackles, fouls, cards, scorers).
    • 4 custom ES|QL tools (get_live_match_stats, get_team_form, get_team_stats_2026, get_upcoming_fixtures) created in Elastic Agent Builder and exposed via a Model Context Protocol (MCP) server.
    • Two execution paths: Elastic Inference Service (EIS / AWS Bedrock) in Agent Builder and an interactive Streamlit chat application powered by Fireworks AI (glm-5p2) with real-time MCP tool invocation visualization.
    • Project summary & architecture documentation.
  • OrthoStream AI: Multimodal Orthodontic QA Diagnostic Agent & Aligner Simulator: Built for the Google I/O Build with AI Hackathon x Google Cloud Labs, OrthoStream AI is an interactive full-stack clinical validation assistant and simulation workspace for clear-aligner treatment planning. It bridges spoken clinical feedback and digital aligner manufacturing by translating doctor voice transcripts into mathematically validated tooth-movement directives with real-time holographic simulation.
    • Powered by the Google GenAI SDK (@google/genai) using Gemini 3.5 Flash for multimodal audio command transcription, natural language clinical intent mapping, and JSON directive generation.
    • Interleaved diagnostic execution flow powered by a Server-Sent Events (SSE) streaming engine delivering dynamic 3D mesh comparisons, stage focus scrubbers, geometric discrepancy cards, and clinical assertion reports.
    • Dual graphics simulation engine featuring interactive HTML5 vector canvases with matrix rotation, tooth-contour morphing, drag-and-drop smile calibration, holographic 3D X-ray bone scanners, and color-coded stress curves.
    • Built with React 19, Tailwind CSS, motion, Node.js/Express, containerized and packaged for Google Cloud Run.
    • Documentation and architecture repository.
    • 3rd place podium at the Google I/O Build with AI Hackathon x Google Cloud Labs
  • CivicAurAI: Created for the Build and solve with AI: San Francisco edition pre-RSAC Hackathon (Day 1 and Day 2), CivicAurAI is a civic infrastructure intelligence and municipal reporting ecosystem connecting citizens, city agencies, and autonomous vehicle fleets.
    • Continuous dashcam and mobile video recording (CivicAurAIRecorder) with synchronized GPS telemetry automatically flags municipal hazards like potholes, illegal dumping, and encampments.
    • Multimodal video processing pipeline built on RocketRide AI and Google multimodal models for automated hazard detection and visual chain-of-custody verification.
    • Spanner Property Graph (CivicGraph) and Spanner RDBMS architecture for temporal episode tracking (Graphity temporal modeling), deduplicating independent AI vehicle detections with human citizen reports, and dynamic service routing to municipal organizations based on geographic district boundaries.
    • Community upvoting mechanics to assess community priority without artificial vehicle telemetry inflation.
    • Civic-AurAI organization and Final submission codebase.
    • Winner of "Best use of Gemini" and RocketRide special prizes
  • Eephor: Oversight Agent for Multi-Agent Systems: Developed during the Reinforcement Learning Hackathon by Cerebral Valley, Eephor is an end-to-end training and evaluation pipeline for specialized LLM oversight agents capable of distinguishing benevolent agents from deceptive ones in complex multi-agent environments like MoltBook.
    • Transforms conversational logs (Ethos Academy) into structured ground-truth datasets across 12 traits (Integrity, Logic, Empathy) with strict thread-level data splitting to prevent contextual data leakage.
    • Estimates hidden internal intent and agent architecture (replicating dynamic SOUL.md profiles) using behavior trait rubrics and few-shot multi-turn conversational reasoning.
    • Two-stage training pipeline: Parameter-Efficient Fine-Tuning (PEFT) with LoRA via Unsloth for Supervised Fine-Tuning (SFT) pre-warming, followed by Reinforcement Learning with Verifiable Rewards (RLVR) via Group Relative Policy Optimization (GRPO) in OpenEnv to prevent reward hacking.
    • Published open oversight models on Hugging Face: moltbook-oversight-llama31-1b and moltbook-oversight-llama-3.2-3b-instruct (standard and GGUF quantizations).
    • Eephor Jupyter PyTorch training environment.
  • BarackOllama & BarackOllamaCLI: Built for the Llama Lounge Snowflake Agentic Hackathon, BarackOllama is an agentic resume, GitHub portfolio, and cover letter customization suite designed to tailor candidate applications specifically to targeted job listings.
    • Powered by Google's Agent Development Kit (ADK) and Gemini models for deep resume extraction, code repository inspection, and alignment scoring.
    • Integrates Snowflake for candidate data management and CrewAI Enterprise flows to coordinate research and writing agent teams.
    • Provides both an interactive command-line interface (BarackOllamaCLI) and a full-stack web application with FastAPI server backend and modern frontend.
  • AI-Q Research Assistant with Universal Deep Research (Research-as-a-Code): Developed for the AWS & NVIDIA Agentic AI Unleashed Hackathon 2025, this project synthesizes NVIDIA AI-Q Research Assistant with NVIDIA Universal Deep Research (UDR) into a two-level hierarchical agent architecture for complex, multi-domain research tasks.
    • Two-level agentic system: Level 1 orchestrator (LangGraph) plans research strategy, and Level 2 UDR engine dynamically generates and executes custom Python research pipelines when complex synthesis or comparative analysis is required.
    • Built with NVIDIA NIM microservices: Nemotron-Super-49B NIM for reasoning, reflection, and planning; Llama-3.3-70B NIM for report synthesis; NeMo Retriever NIM for dense vector search; and Docling for semantic document ingestion.
    • Production cloud infrastructure on AWS EKS utilizing Karpenter for dynamic GPU auto-scaling and GPU Operator provisioning.
    • Interactive React/Next.js frontend featuring CopilotKit for real-time streaming visualization of the agent's thought process, tool invocations, and synthesized citations.
  • ¡alacard!: The Self-Improving Cookbook for AI Agents: Created during WeaveHacks 2: Self-Improving Agents Hackathon hosted by Weights & Biases and Google Cloud, ¡alacard! is an open-source instruction layer and agentic factory that composes, executes, traces, and autonomously repairs Jupyter notebooks.
    • Users define high-level JSON "Recipes" selecting models (Gemini, Llama) and tools (Tavily search, Browserbase Stagehand web automation, Mastra integrations).
    • Multi-agent orchestration powered by Google ADK: a "Chef" agent writes full runnable notebooks, which execute safely inside isolated Daytona sandboxes.
    • End-to-end observability and State-Action-Reward (SAR) logging powered by W&B Weave tracing to capture every tool call, model prompt, and intermediate output.
    • Autonomous self-healing loop: when a code cell fails, an automated "Fixer" agent inspects the Weave execution trace, queries Vertex AI / W&B Inference for code patches, applies updates, and re-executes until passing.
    • Validated notebooks are cataloged in a Supabase-backed hub with CopilotKit AG-UI streaming frontends for community sharing and remixing.
    • ¡alacard! Enterprise and landing page.
    • Finalist (top 6 projects) at the WeaveHacks 2: Self-Improving Agents Hackathon
  • Stigmergy ODEA (Open Distributed Edge Agents): Built for the Google Gemma 3n Impact Challenge on Kaggle, Stigmergy ODEA is a decentralized, resilient multi-agent swarm intelligence system designed for tactical edge operations and first responders in post-disaster environments (earthquakes, wildfires, floods).
    • Deployed on fleets of resource-constrained Android edge devices (drones) running on-device Google Gemma 3B multimodal models.
    • Uses Android Nearby Connections to establish a peer-to-peer, low-power mesh network operating completely independently of cellular or cloud infrastructure.
    • Implements dynamic Commander/Subordinate swarm hierarchies with automatic failover and leader re-election protocols if a commander unit is lost or compromised.
    • Built-in defenses against adversarial prompt injection and device impersonation attacks, combined with multimodal perception to reconcile visual camera feeds against expected mission targets.
    • Built upon and extended from the official Google AI Edge Gallery demo app.
    • Open Distributed Edge Agents organization.
  • Public Health Application: Developed during the AI Expo SCSP Hackathon 2025 (hosted by Booz Allen Hamilton and the Special Competitive Studies Project), this project is a public health surveillance, monitoring, and analysis system.
    • Intelligent multi-agent analytical system (PublicHealthDashboardAgent) built with LangGraph to ingest live health indicators, analyze risk trends, and automatically generate executive health dashboards.
    • Multi-LLM support across OpenAI GPT-4 and Anthropic Claude for synthesizing complex epidemiology data and generating situational summaries.
    • FastMCP (Model Context Protocol) server providing standardized, real-time health data tools (alerts retrieval, trend forecasting, emergency summaries) via Server-Sent Events (SSE) transport.
    • Full-stack architecture backed by FastAPI with async API endpoints and VS Code development task workflows.
  • LlamaLink (LlamaLinked): Built during the LlamaCon Hackathon hosted by Meta & Cerebral Valley, LlamaLink is a low-pressure, WhatsApp-first professional networking platform with an automated weekly matchmaking cadence.
    • Integrates LinkedIn profile scraping via Gumloop/Firecrawl, Supabase (PostgreSQL + pgvector) for profile storage, Meta Llama 4 Maverick (17B on Groq) for summarization, Mailgun for mutual warm introductions, and automated post-meeting active-listening coaching loops.
    • Contributed core AI engineering components:
      • Synthetic Data Generator: Document parsing, synthetic QA and Chain-of-Thought (CoT) reasoning pair generation via synthetic-data-kit, quality filtering, and Alpaca/HuggingFace dataset preparation for fine-tuning.
      • Prompt Optimization Tool: Streamlit application utilizing llama-prompt-ops to iteratively evaluate and optimize prompts against input-output training pairs.
      • System prompt engineering, guardrail implementation, and product architecture specifications (PRD repository and Web frontend).
  • Open Multi-Modal Personal Assistant (OpenMMPA): Developed for the Gemini API Developer Competition, OpenMMPA is an open-source, multi-modal personal AI companion application designed to turn smartphones and smartwatches into intelligent personal assistants without requiring dedicated AI pins.
    • Cross-platform mobile architecture built with Flutter (iOS & Android) and powered by Google Gemini 1.5 Flash and Pro models (via Vertex AI / Firebase Vertex AI).
    • On-device local Vector Database providing persistent conversation history and personalized on-device Retrieval Augmented Generation (RAG).
    • Hands-free multimodal interaction supporting camera visual exploration, native Android STT/TTS, and Google Chirp speech recognition for real-time translation and voice chat.
    • Tool and function calling capabilities including location-aware queries, real-time weather forecasting, Tavily web search, and currency exchange rates.
    • OpenMMPA Demo Video and DeepWiki Architecture Documentation.
  • Eco Trivia Quiz Application (with chat agent on the side): Eco Trivia is a web / mobile game about climate change. Apathy and information overload prevent people getting involved with climate action. I gathered Google Developer Student Club (GDSC) Fresno members Zheng Wei Ng and Ran Hao Wong to create a quiz game about sustainability. We were exploring how far we can go with low-code / no-code tools:
  • Eco Oracle Chat Agent: A no-code / low-code agent specialized in Sustainable Development Goals via RAG (Retrieval Augmented Generation). It uses the Google Cloud Platforms Agent Builder offerings which are Generative AI successor to DialogFlow CX conversational agents, and it offers no code "playbooks" (or behavior scripts), managed RAG, function calls and powered by Google Gemini under the hood.
  • LLM Knowledge agent with RAG Fusion: Using Cohere chat, document guided and web connector guided agents, Weaviate services and LangChain, I implemented a RAG Fusion solution with a Streamlit UI. I also experimented with Cohere's Command R large language model.
  • QBRAG (QnA Boosted RAG) Knowledge LLM agent on Vectara platform: A company's knowledge bases often times don't answer the wide variety of questions a user could come up with. A Customer Support system ideally could answer specific (but wide variety) questions about the company's systems and knowledge (example: "How can I enter Cash Flow in ThruThink?"). But sometimes the user asks generic questions, such as "What is Cash Flow?" which could be sourced from the mind of a giant LLM model and / or the internet. My submission is essentially synthetic data enrichment of the knowledge base by reusing generated questions and answers. To help and boost the performance by leveraging Question and Answer generation techniques - normally used for fine tuning but in this case - for knowledge base augmentation / indexing enrichment. The generated questions could support specific user queries potentially better matching than a "non focused" indexed generic knowledge base.
  • GeMMAA (Generative Multi Modal Augmented Reality Art) application by AIWeavers for Multi Modal hackathon by Weaviate at AIx Innovation Summit. Team repository: https://github.com/AIWeavers/GeMMAA. An AR mobile application which can generate 3D art from a text prompt and also music from a text prompt with the option of minting an NFT from the artwork. We use 8thwall for the front-end, NFTPort for NFT minting, GCP cloud functions for back-end, Meshy for 3D model generation, Meta's music generator, and most importantly Weaviate's multi-model embedding to allow search and organization of the artwork. The team:
  • Knowledge LLM agent with RAG utilizing low-code / no-code cnvrg LLMaaS platform: At the time of the project (Intel Innovation Bridge hackathon @ Intel Innovation 2023) cnvrg's LLMaaS (LLM as a service) by Intel was in alpha stage. The LLMaaS service provided a no code environment to fine tune and customize agent solutions with ease. Due to the development stage of this novel product I could use two toy datasets to fine tune an LLM, and I applied RAG (Retrieval Augmented Generation) with the help of a PineCone vector database, using ChatGPT's embeddings API. I also got familiar with GRadio and similar front-ends. Due to confidentiality the materials I can share are limited.
    • Winner of the Intel Innovation Bridge hackathon's LLM category @ Intel Innovation 2023
  • Gen AI Rock Star Challenge: Song "Xeon Fury" of the band "Nerdcore Annihilation". Participants were tasked with generating a poster or album cover for an imaginary rock band, along with generating the lyrics of an imaginary song for said band. We were provided with a Gaudi2 accelerated deep learning AWS EC2 instance equipped with a Runway 1.5 text-to-image stable diffusion model, and also a Vicuna 7-based LLM model. We increased the steps of the stable diffusion model from 10 to 50 to conclude with much better images. The EC2 box was inaccessible for a while, and in the meantime, we experimented with some other LLM models for lyrics, such as ChatGPT and PaLM, as well as the officially accessible Runway Gen 1. We used the LLM to get the band name (Nerdcore Annihilation) and the song name (Xeon Fury) as well. The stable diffusion model couldn't feature text on the image precisely, so we used Runway's web-accessible edit feature to add an old school Intel Inside logo and modify the center image of the front cover. The back cover was purely generated by the Gaudi2 instance (notice how the wafer is a pastry wafer and not a chip wafer). We adjusted the prompt many times. I used my geek knowledge to provide desired keywords for the lyrics: MMX, AVX, AVX2, AVX512, AMX, Xeon Phi, Gaudi 2, Optimum Habana, 288-core Xeon (freshly announced). We mixed the best parts of the various tries together to create a hilarious final set of lyrics.
    • Winner of Intel GenAI Rock Star Challenge @ Intel Innovation 2023
  • Multi Modal Generative AI Travel blog helper app: An app indexes pictures taken during travel trips or everyday experiences. With the power of multi modal embedding models and vector databases it is able to categorize them, describe them and help document the experiences. The back-end uses Chooch for image to text tasks and Google Cloud Platform for everything else such as asset management, search, summarization and RAG. The front-end is written in Kotlin and uses Compose UI by Kevin Moore GDE. The project management, iteration on ideation, user interviews, feasibility, presentation of our work was done by Yiru and Coco He.
  • CSU Fresno Recycling Trashcan locations - Outdoor Augmented Reality Map leveraging Geospatial API and Terrain Anchors to guide students to recycling trashcans
  • CSU Fresno Tree Walk - Outdoor Augmented Reality Map leveraging Geospatial API and Terrain Anchors to guide visitors along the Tree Walk, boost engagement gamification as well as advanced Generative AI method: segmentation for virtual watering
  • AMP personal website - Google Registry Website Competition Winner
  • Deal-O-Round - Flutter based reimplementation of a PlayN-based multi-platform (Android / iOS / web) game

More projects worth mentioning:

Blog Posts

Github Stats
Github Streak
Top Language

About

GitHub profile repository

Resources

Readme
Activity

Stars

0 stars

Watchers

1 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors