Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

OpenAI's Assistants API

OpenAI's Assistants API simplifies AI integration for developers, eliminating the need for managing conversation histories and providing access to tools like Code Interpreter and Retrieval. The API also allows developers to integrate their own tools, making it a versatile platform for AI assistant development.

General
AuthorOpenAI
DocumentationLink
TypeAI Assistant

Model Overview

The Assistants API enables developers to create AI assistants using OpenAI models and tools. It supports various functionalities such as managing conversation threads, triggering responses, and integrating customized tools.

Assistants API Tutorials

Technology Resources

The Assistants API allows developers to construct AI assistants within their applications. An assistant can leverage models, tools, and knowledge to respond to user queries effectively. Presently supporting Code Interpreter, Retrieval, and Function calling, the API aims to introduce more tools developed by OpenAI while also allowing user-provided tools on the platform.

To explore its capabilities, developers can use the Assistants Playground or follow the integration guide in the official documentation. The integration process involves defining an Assistant, enabling tools, managing conversation threads, and triggering responses.

Edit on GitHub

OpenAI Assistants API AI technology Hackathon projects

Discover innovative solutions crafted with OpenAI Assistants API AI technology, developed by our community members during our engaging hackathons.

Central Finite AI

Central Finite AI

Central Finite AI is an enterprise AI operating system built to bring LLM-assisted automation to IT operations and core-banking engineering — two domains where mistakes are expensive and blind trust in AI output is not an option. The system runs on Google Cloud Run with a FastAPI backend, a React NOC-style dashboard, and ChromaDB as a persistent knowledge base that ingests T24 (Temenos Transact) documentation, sample routines, and prior incident history. On top of that sits a five-agent investigation pipeline — Knowledge, Graph, Connector, Investigation, and DevOps agents — that traces incidents through a dependency graph using breadth-first impact analysis, pulls context from read-only Git/Jira/SQL connectors, and proposes remediation. Two features push this from "chat about your systems" into "act on your systems," safely: - An OFS Builder that constructs T24 OFS messages (the format T24 uses for programmatic transaction submission) either from a deterministic manual form or from a plain-language description via Fireworks-hosted inference. Every generated message is shown in an editable preview before submission. - A Routine Builder that drafts T24 subroutine source code (jBC/TAFJ) grounded in ingested T24 documentation, flagging any guessed field names in review notes rather than silently inventing them. Nothing compiles, attaches, or submits to a live T24 instance automatically. Every action — writing a routine, compiling it, filing a PGM.FILE entry, attaching it, or submitting an OFS message — passes through a deny-by-default execution engine with an audit-logged approval step in the dashboard's Actions tab. A local Windows/TAFJ compile agent polls the Cloud Run backend and only executes actions once a human has approved them. The result is a system that gives IT operations and T24 developers LLM-speed drafting and investigation, without giving an LLM unsupervised write access to production banking infrastructure.

RouteZero — Autonomous Incident Routing

RouteZero — Autonomous Incident Routing

RouteZero is a four-agent autonomous incident routing system built for engineering teams drowning in manual triage. When a production incident occurs, engineers paste a stack trace and RouteZero handles the rest. Agent 1 (Classifier) uses zero LLM calls — pure deterministic regex and keyword scoring — to classify the failure type, detect the affected service, and assess blast radius. Every decision is fully auditable. Agent 2 (Router) applies rules-first priority logic: critical path plus production automatically escalates to P1, SLA breaches under 60 minutes trigger P1, and deployment timing enables probable cause detection. LLM is only consulted when confidence falls below 65%. Agent 3 (Ticket Writer) uses Fireworks AI with Gemma to assemble role-appropriate content for three stakeholders simultaneously — a full technical ticket for the engineer, a leadership-framed summary for the team lead, and a five-sentence plain English digest for the manager. A hallucination validator rejects any number or filename in AI output not present in the verified input facts. Agent 4 (Architectural Auditor) mines incident history for recurring patterns — same file and line number across multiple incidents, service stress signatures, and cascading failures — then files proactive PLM tickets before the next incident happens. A human approval gate sits between routing and action. Nothing is created or sent until the engineer clicks approve. Built on FastAPI, Streamlit, DuckDB, and Fireworks AI.

Adaptive General-Purpose AI Agent

Adaptive General-Purpose AI Agent

`Adaptive General-Purpose AI Agent is a Track 1 submission designed to solve hidden natural-language tasks efficiently across eight categories: factual question answering, mathematical reasoning, sentiment classification, summarization, named entity recognition, code debugging, logical reasoning, and code generation. The system is built as a containerized batch-processing agent that reads tasks from /input/tasks.json, classifies each prompt, selects an appropriate solving strategy, writes answers to /output/results.json, and exits automatically in the required competition format. The core idea is to avoid using a large external model for every task. Instead, the agent applies a hybrid routing architecture. Simple structured tasks such as arithmetic word problems, exact one-sentence summaries, basic sentiment classification, and selected logic puzzles are solved locally using deterministic logic. More open-ended or complex tasks such as factual knowledge questions, named entity extraction, code debugging, and code generation are escalated to an allowed Fireworks model. This approach improves token efficiency while preserving broad task coverage. The project also includes metrics tracking for route selection, Fireworks usage, latency, and token consumption. In a validated benchmark sample, the hybrid strategy reduced Fireworks calls from 6 to 4 and reduced total Fireworks token usage from 2011 to 1459, achieving a 27.4% token reduction. This demonstrates that selective escalation can meaningfully reduce external model usage without breaking the submission workflow. The system is modular, competition-aligned, and designed for further improvement through stronger local solvers, better routing confidence, and more efficient output control.`

AITinerary

AITinerary

AITinerary – Your AI Travel Co-Pilot AITinerary is an AI-powered travel planning platform designed to simplify every stage of a trip—from discovering destinations to creating personalized itineraries and exploring hidden gems. Instead of spending hours researching across multiple websites, users simply describe their travel preferences, budget, trip duration, and interests, and AITinerary generates a complete travel plan tailored to them. One of the core ideas behind AITinerary is bridging the gap between travel inspiration and actual trip planning. Today, many people discover amazing destinations, restaurants, and experiences through Instagram Reels and YouTube videos, but planning a trip around that content is still a manual process. AITinerary aims to let users provide a Reel or YouTube link and transform that inspiration into a practical itinerary with recommended attractions, restaurants, accommodations, transportation, and nearby experiences. The platform also acts as an intelligent travel companion throughout the journey. It recommends hidden gems beyond popular tourist attractions not to miss, adapts plans based on user preferences, provides contextual information about places, helps optimize travel budgets, and enables expense tracking and bill splitting for groups. For the MVP, the focus is on AI-generated itineraries, social media-inspired trip planning, personalized recommendations, and intelligent travel assistance. The architecture is designed to integrate with travel providers and booking platforms in the future, allowing users to seamlessly transition from planning to booking within a single experience all at one place. By combining generative AI, travel data, and personalization, AITinerary aims to become an all-in-one travel assistant that helps users spend less time planning and more time experiencing memorable journeys. "From inspiration to itinerary in seconds. See it. Plan it. Experience it."

AMD Hybrid Token-Efficient AI Agent

AMD Hybrid Token-Efficient AI Agent

This project implements a hybrid token-efficient AI agent designed to solve a wide variety of tasks while minimizing unnecessary LLM usage. The system first classifies each incoming task and routes it to the most appropriate execution path. Deterministic problems such as mathematical computations, fuzzy matching, symbolic reasoning, and rule-based operations are solved locally for maximum speed and zero token cost. More complex tasks including summarization, code generation, logical reasoning, debugging, named entity recognition, and factual question answering are handled using Fireworks AI models selected dynamically through the provided ALLOWED_MODELS environment variable. The application is fully containerized using Docker, reads tasks from /input/tasks.json, writes results to /output/results.json, and complies with the evaluation harness by using environment-based configuration, Linux AMD64 compatibility, and efficient routing to reduce latency and inference costs while maintaining high accuracy.

humongous kitten returns

humongous kitten returns

StyleCap is a video captioning agent with one uncompromising goal: describe what actually happens in a clip, and say it in exactly the voice you ask for. Hand it any short video — a busy city street, a kitten in a garden, a chef plating a dish, a sudden downpour, a decisive match point — and it returns four captions in four distinct registers: a clear, factual account; a dry, sarcastic take; a witty, tech-flavored line; and warm everyday humor. Styled captioning hides a trap: tone and truth tend to fight. The instant a caption reaches for a joke, it usually stops describing the video. StyleCap is built so the two never trade off — every caption stays firmly anchored to what is really on screen, so someone who never saw the clip still learns what happened, while each style lands unmistakably. It holds up across nature, people, animals, sport, food, weather, and technology, and stays fast, reliable, and consistent on footage it has never seen. Accurate first, stylish always.