Agronomist

Smart resource allocator

VolunteerConnect is a comprehensive volunteer coordination platform designed to solve the operational challenges faced by NGOs and volunteer organizations. Built with a modern serverless, event-driven architecture, the platform brings together NGO administrators, coordinators, and volunteers into a unified digital ecosystem. At its core, VolunteerConnect leverages Firebase (Firestore, Cloud Functions, FCM) for real-time data sync and push notifications, ensuring all stakeholders stay instantly informed about task updates, event changes, and volunteer assignments. The platform integrates Gemini 1.5 Flash AI to intelligently match volunteers to opportunities based on their skills, availability, and proximity — reducing the manual overhead for coordinators. The mobile-first experience, built with React Native Expo for Android, allows volunteers to discover opportunities, check in to events, log hours, and communicate on the go. Meanwhile, coordinators and admins access a powerful Next.js 14 web dashboard (hosted on Vercel) for managing events, reviewing applications, generating reports via Google Sheets API, and overseeing role-based access across the organization. Google Maps and Distance Matrix APIs power location-aware features, helping volunteers find nearby opportunities and enabling coordinators to plan logistics efficiently. The platform also incorporates a robust RBAC (Role-Based Access Control) system and an immutable audit log, ensuring accountability, transparency, and data security across all operations. VolunteerConnect is built to scale — from small local NGOs to large multi-region organizations — making volunteer management more efficient, data-driven, and impactful.

MedSignal

Doctors in high-volume clinics don't miss critical signals because they lack knowledge — they miss them because they lack time. Patient input arrives fragmented, informal, and mixed-language. Existing AI tools assume clean records and hallucinate when data is missing. MedSignal is built for that reality. It accepts raw, unstructured patient text and produces a severity-ranked, uncertainty-aware clinical report in under 2 seconds. Missing data is flagged explicitly, never silently filled in. Five CrewAI agents run on AMD Instinct MI300X GPUs. The Intake Agent structures raw text into validated clinical JSON without assuming anything not present in the input. Three agents then run in parallel: the DDx Agent generates a ranked differential diagnosis weighted by severity and probability; the Red Flag Agent runs 28 deterministic clinical rules, live OpenFDA drug interaction lookups, and bounded LLM reasoning; the Consistency Agent detects contradictions in patient history. The Summary Agent assembles the final prioritized report. Parallel execution on MI300X cuts wall time from ~4.5s to ~2s. Agents modulate each other's confidence. Contradictory history lowers LLM weights, adds an explicit warning, and escalates specific risks — cross-agent reasoning that makes the system behave like a clinical team, not a pipeline. The rule engine covers India-specific emergencies Western clinical AI routinely misses: snake envenomation, dengue warning signs, organophosphate poisoning, scorpion envenomation, tetanus-prone wounds, and paracetamol-alcohol hepatotoxicity. Severity is always computed deterministically — never trusted to the LLM. An ALWAYS_CRITICAL guard prevents true emergencies from being downgraded. Built for the AMD Developer Hackathon 2026. FastAPI backend, React/Vite frontend with real-time SSE streaming, deployed on HuggingFace Spaces and Vercel.

AndesOps AI: Multi-Agent Tender Intelligence

Overview Public procurement in Chile (Mercado Público) involves thousands of documents with complex legal and technical requirements. For most companies, analyzing these opportunities is slow, risky, and expensive. AndesOps AI solves this by deploying a "Virtual Board of Experts"—an agentic workflow that processes tenders in seconds. The Solution: The "Expert Round Table" Unlike simple RAG systems, AndesOps AI uses an Agentic Orchestration Layer where specialized AI agents collaborate: ⚖️ Legal & Compliance Agent: Scans administrative bases for critical deadlines and compliance gaps. 🏗️ Technical Architect Agent: Evaluates the feasibility of requirements against the company's stack and experience. 📊 Strategy & ROI Agent: Calculates potential profitability and competitive risks. 🧠 Orchestrator: Consolidates findings into a Strategic Fit Score (0-100) and a professional report. Technical Stack & AMD Integration Backend: FastAPI (Python) with asynchronous parallel agent execution. Frontend: Next.js 14 + Tailwind CSS (Enterprise-grade UI). AI Engine: Multi-agent system designed to run high-performance inference. AMD Advantage: Optimized for AMD Instinct™ GPUs using ROCm. By leveraging AMD's massive memory bandwidth, AndesOps AI can process hundreds of pages of bid documents simultaneously, providing near-instant feedback for time-sensitive public bids. Business Value 90% Reduction in manual analysis time. Risk Mitigation: Early detection of "hidden" legal requirements that could lead to disqualification. Increased Win Rate: AI-driven proposal drafting that highlights the company's competitive advantages. 4. Technology & Category Tags (Etiquetas) AI Agents Agentic Workflows ROCm AMD Developer Cloud FastAPI Next.js GovTech Llama-3 5. Submission Checklist (Criterios de Evaluación) Application of Tech: Mention that the project is built to leverage ROCm for high-throughput document processing. Originality: Emphasize the "Expert Panel" approach over a single-agentchatbot

NyayaLLM: Indian Criminal Law AI

Navigating the recently overhauled Indian criminal justice system, which introduces the Bharatiya Nyaya Sanhita (BNS), Bharatiya Nagarik Suraksha Sanhita (BNSS), and Bharatiya Sakshya Adhiniyam (BSA), is a massive challenge for both legal professionals and everyday citizens. NyayaLLM solves this by providing a highly accurate, domain-specific AI assistant trained exclusively on these new legal frameworks. We leveraged the immense compute power of AMD MI300X instances to fine-tune a Llama 3.1 model, ensuring rapid iteration and high throughput during the training phase. By utilizing ROCm and Hugging Face libraries, we adapted the base model to understand complex legal jargon, procedural nuances, and the specific relationships between the old IPC/CrPC sections and the newly implemented BNS/BNSS codes. The resulting application allows users to query specific legal scenarios, understand compliance requirements, and draft preliminary legal summaries with significantly reduced hallucinations compared to standard, general-purpose models. Our ultimate goal with NyayaLLM is to democratize legal knowledge, increase productivity for law firms dealing with the transition, and ensure that the complexities of the new Indian criminal laws do not become a barrier to justice.