Adriel – LLM, Python, AWS

AI-powered logistics monitoring platform that helps security analysts track truck shipments, investigate sensor-based alerts, and coordinate response actions. The solution integrates data from IoT tracking devices to centralize alert management, shipment visibility, analyst workflows, and escalation processes. LLM-powered comment analysis automatically extracts structured context from analyst notes, improving information visibility and reducing response times by 20%.

• Designed and built the entire application end-to-end, including frontend, backend, database architecture, and deployment;
• Set up the data layer in Supabase (PostgreSQL), including tables for alerts, alert comments, alert status tracking, manual status overrides, and reporting views, and implemented database triggers for automatic alert resolution;
• Integrated the LLM processing pipeline to extract structured information from analyst comments, serving as the core differentiator of the product;
• Built the IoT partner integration through webhooks and scheduled synchronization jobs to keep alerts, shipments, and statuses up to date;
• Owned application deployment through Lovable and managed the staging environment;
• Produced detailed technical documentation of the system architecture and investigated and resolved data inconsistencies, including comment display and billing-related issues;
• Drove iterative product improvements, including server-side pagination, time zone handling, status inference logic, manual override workflows, and the Shipments without Alerts view, based on analyst feedback;
• Collaborated with stakeholders on the product roadmap, tracking deliverables related to workflow replacement initiatives, shipment-alert matching, AI-powered analysis, observability, and automated testing.

AI-powered logistics operations automation platform designed to monitor delivery workflows through a third-party logistics system and surface only actionable exceptions to human analysts. The solution combined Computer-Use Agents (CUAs), browser automation, and multimodal data processing to track delivery milestones, interpret platform activity, and automate routine monitoring tasks. By shifting analysts' focus from status checking to exception resolution, the platform increased operational productivity by 4x.

• Designed the end-to-end agent architecture, selecting and orchestrating multiple CUA providers and defining their interaction with the browser environment through Browserbase and Playwright;
• Built the multimodal extraction layer that converted HTML content and screenshots from the third-party platform into structured delivery status data;
• Implemented exception-handling logic to distinguish routine delivery updates from issues requiring human intervention;
• Built and tested the process-stops API that powered the workflow, including validation of delivery outcome edge cases such as late, early, partially delivered, and damaged deliveries;
• Measured solution impact, contributing to a 4x increase in logistics analyst productivity by reducing manual monitoring of routine delivery statuses.

RAG-powered knowledge assistant designed to help logistics support teams quickly access information stored across extensive Standard Operating Procedure (SOP) documentation. The solution combined semantic search, multi-agent orchestration, and large language models to deliver accurate, document-grounded answers to operational questions through a conversational interface. By eliminating manual document searches, the platform reduced customer support response times by 50%.

• Architected the multi-agent system that determined when to operate as a general conversational assistant versus when to perform retrieval-augmented generation using SOP documentation;
• Integrated AWS Bedrock as the underlying LLM platform for response generation;
• Set up and maintained OpenSearch as the vector database for document indexing, embedding storage, and semantic retrieval;
• Implemented CrewAI to orchestrate agent interactions and query-routing workflows;
• Connected the SOP document repository to the retrieval pipeline, enabling accurate context retrieval and document-grounded responses.

AWS-native video intelligence platform designed to transform large volumes of unstructured video content into searchable, machine-understandable assets. The solution supported multiple use cases, including media archives, stock footage libraries, and live sports broadcasts, combining multimodal AI, semantic search, and event detection to enable natural language discovery and automated content analysis. The platform delivered 95% search accuracy for curated video libraries, enabled real-time search across multi-thousand-hour archives, and improved sports event detection performance by 5x.

• Designed and owned the end-to-end architecture of a modular video intelligence platform, including batch and real-time ingestion pipelines, multimodal enrichment workflows, hybrid search infrastructure, and natural language query interfaces;
• Built the video annotation pipeline using parallel AI processing for video understanding, audio transcription, and object tracking, merging outputs into unified searchable metadata;
• Designed and implemented chunking strategies with configurable processing parameters to balance cost, latency, and annotation accuracy across different video workloads;
• Integrated object-tracking capabilities for real-time sports analysis and event detection workflows;
• Built and optimized OpenSearch-based hybrid search capabilities combining semantic vector search with structured filtering;
• Developed natural language search agents to support fast retrieval across large-scale video archives;
• Contributed to platform deployment and infrastructure automation using AWS-native services, serverless architecture, and infrastructure-as-code tooling;
• Evaluated and refined model selection, prompt strategies, and retrieval workflows to improve production readiness and search quality across multiple client use cases;
• Delivered measurable business outcomes, including 95% semantic search accuracy, a 5x improvement in sports event detection performance, and real-time search capabilities across multi-thousand-hour video archives.

AI-powered web data intelligence platform designed to aggregate, organize, and search business information collected from across the web. The solution combined semantic retrieval, website clustering, and machine learning models to identify when different web sources referred to the same business or topic, enabling accurate search, brand monitoring, market research, and business intelligence workflows. The platform became the core search engine powering multiple products and contributed to a 30% increase in company revenue.

• Owned the end-to-end machine learning pipeline, including feature engineering, model development, training, deployment, monitoring, and inference optimization;
• Researched and developed machine learning approaches for website clustering and similarity detection using transformer-based language models;
• Fine-tuned language models for pairwise website comparison and classification tasks using real-world web data and synthetic training datasets;
• Built retrieval-augmented workflows that combined vector search and machine learning models to identify and group related websites;
• Designed and optimized ElasticSearch and OpenSearch retrieval architectures, including schemas, indexing strategies, and hybrid search workflows;
• Implemented hybrid search capabilities combining semantic embeddings with keyword-based retrieval techniques to improve candidate selection quality;
• Tuned retrieval and inference pipelines to provide highly relevant inputs for downstream machine learning models;
• Architected and maintained the MLOps infrastructure, including model packaging, deployment, scalable inference endpoints, and cloud-based training workflows;
• Implemented parallelization and performance optimization strategies for model inference and retrieval operations to improve scalability and responsiveness;
• Collaborated directly with clients to understand business requirements and align machine learning solutions with product needs.