Summary

Commercial ASR systems fail badly for Myanmar (Burmese), producing a Word Error Rate (WER) of >80% in zero-shot scenarios. This is due to the complex, non-segmented script, tonal distinctions, and an extreme lack of labeled conversational datasets. Engineered a robust ML pipeline to clean, filter, and augment (speed perturbation) a curated 54.2-hour Myanmar speech dataset from FLEURS, OpenSLR, and YODAS. Fine-tuned three multilingual transformer models (Whisper v3 Turbo, Dolphin Whisper-v2, SeamlessM4T v2 Large) using PyTorch and HuggingFace Accelerate on Vast.ai RTX 4090 GPUs. Established a full MLOps tracking stack using MLflow, MinIO, Label Studio, and Argilla. Handling severe hardware constraints, mitigating catastrophic forgetting of pre-trained audio features, and validating audio quality to filter out over 30% of raw corrupted data. Implemented a frozen-encoder training strategy to stabilize parameter updates. Achieved a dramatic 67 percentage point reduction in WER (down to 33.02% with Dolphin) and a 75 percentage point drop in Character Error Rate (CER to 13.04% with SeamlessM4T). Successfully packaged the research into an interactive Next.js presentation and Streamlit inference demo.

Click card to read full summary

Summary

Healthcare and pharmaceutical e-commerce platforms require absolute real-time synchronization between digital storefronts, multi-vendor warehouses, procurements, and brick-and-mortar retail point-of-sale (POS) systems to prevent critical stock discrepancies, while ensuring comprehensive auditability and fault-tolerant distributed transaction safety. Architected an event-driven microservices ecosystem using NestJS, running a database-per-service architecture with Postgres 17 and Prisma ORM. Engineered high-speed synchronous transport using gRPC, and asynchronous pub-sub event dispatching via Kafka/Redpanda for domain events, CDC, and sagas. Integrated RabbitMQ for resilient notification task queues, MeiliSearch for sub-second catalog search projection workers, and an enterprise observability suite consisting of OpenTelemetry, Tempo, Loki, Prometheus, and Grafana. Ensuring transactional consistency across distributed boundary contexts (IAM, Catalog, Order, Notification, Inventory, Payment) during checkout sequences, preventing out-of-order Kafka message consumption, and setting up centralized distributed tracing and log correlation across asynchronous event streams. Successfully engineered a fault-tolerant Checkout Saga pattern, eliminated transaction deadlocks, reduced inventory synchronization latency to near-zero, achieved real-time product search indexing with MeiliSearch projection workers, and established full log-metric-trace correlation via OpenTelemetry and Grafana.

Click card to read full summary

Summary

Deploying a highly responsive and secure AI chatbot capable of advanced document reasoning and URL threat classification, while maintaining strict data privacy through local, on-premise LLM inference. Architected a modular FastAPI server integrating local LLM inference via Ollama (DeepSeek-r1:7b). Implemented a robust Retrieval-Augmented Generation (RAG) pipeline utilizing ChromaDB for fast vector search and deployed transformer models for URL classification. Engineered a comprehensive observability stack using Prometheus, Grafana, and Loki for real-time monitoring of custom LLM latency and error metrics. Ensuring low-latency inference with local large language models, configuring seamless Docker multi-container networking (API, Metrics, Vector DB, Log aggregation), and maintaining robust SSL termination using Caddy. Delivered a production-ready, highly observable AI server capable of accurate cyber-threat analysis. Won the APTYPS Champion award for its technical excellence, robust architectural design, and practical security applications.

Click card to read full summary

Summary

During Myanmar flood relief efforts, organizations and volunteers suffered from disjointed communication. This led to redundant supply deliveries to the same locations while other heavily affected areas remained underserved. Developed a cross-platform mobile application using React Native, Expo, and Firebase. Implemented a centralized database for organizations and volunteers, integrated location-based services to identify underserved areas, and utilized push notifications for real-time alerts. Additionally, created an IoT device simulation using Arduino to prototype hardware integration. Ensuring secure user authentication and reliable push notifications across diverse mobile devices. Maintaining real-time data sync and location tracking to accurately analyze and improve resource distribution efficiency during emergency scenarios. Delivered an award-winning mobile application (Build Myanmar Challenge — 2nd Runner Up) that successfully centralized relief coordination, enabling efficient resource tracking and data analytics to optimize supply distribution.

Click card to read full summary

Summary

Scaling highly personalized, real-time recommendations (posts, pages, channels, and friend suggestions) dynamically inside a high-traffic social media application, while maintaining secure API-key authentication, strict rate-limiting, and sub-second latency. Built a modular FastAPI recommendation engine leveraging Neo4j (5.x) for relationship-based social graph queries. Implemented content-based and collaborative filtering algorithms (TF-IDF, cosine similarity, collaborative filtering). Integrated Redis (7.x) for session caching and recommendation buffering, Gunicorn multiprocess workers for concurrency scaling, and a fully automated CI/CD pipeline (Black, isort, Bandit, pytest) with Kubernetes container orchestrations. Handling high Neo4j graph query complexity for nested social-graph traversals, resolving asynchronous event loop conflicts under concurrent Gunicorn workers, and configuring non-root read-only root filesystem security rules inside Kubernetes pods. Delivered microsecond recommendation responses via Redis cache layers, successfully protected API surfaces with rate-limited cryptographically secure API keys, and automated production-ready multi-stage Docker builds to public GHCR registries.

Click card to read full summary

Summary

Building a unified platform that seamlessly combines e-learning, 1-on-1 coaching, event management, and a social community. The application required an intricate backend API to drive complex business logic (e-commerce, progress tracking) alongside a highly responsive frontend for handling rich media and real-time messaging. Developed robust backend REST APIs and core business logic using Laravel, ensuring strict data integrity and secure e-commerce transactions. Adopted a cutting-edge frontend stack utilizing Next.js 15, React 19, Zustand, and SWR for efficient data fetching and optimistic UI updates. Deployed the frontend to Cloudflare using OpenNext. Synchronizing complex state between the Laravel backend APIs and the Next.js frontend, especially for multi-step forms and secure checkouts. Implementing a scalable WebSocket architecture for instant messaging. Ensuring reliable large media file uploads over unstable connections using tus-js-client. Delivered a true full-stack solution with a secure, highly performant Laravel backend powering a fast Next.js edge application on Cloudflare Workers. Integrated robust real-time communication, resilient resumable uploads, and full internationalization (`next-intl`) support from day one.

Click card to read full summary

Summary

Creating a robust foundation for a dual-purpose therapeutic solution: a seamless customer-facing e-commerce/educational platform and a powerful, data-rich administrative backend for business operations. Engineered an API-first architecture using Laravel 12 (PHP 8.3) and a reactive Vue.js 3 frontend. Built a highly customized backend interface utilizing Filament v3 with strict Role-Based Access Control (RBAC). Established comprehensive automated testing pipelines using Pest (PHP) and Vitest (JS). Implementing complex multi-variant product catalogs, ensuring real-time bidirectional updates via Laravel Reverb, and architecting a flexible content management system capable of handling rich media and publishing workflows via Spatie Media Library. Delivered a highly scalable digital health platform featuring real-time business intelligence dashboards (Apex Charts), secure token-based API authentication (Sanctum), and extensive Excel reporting capabilities for administrators.

Click card to read full summary

Summary

Organizations needed a highly secure data entry system capable of tracking fine-grained user actions, IP addresses, and device metadata across the platform while strictly enforcing role-based access controls and streamlining complex administrative workflows. Architected a secure Data Entry ERP application using Laravel 12 and PHP 8.2. Engineered a robust administrative dashboard using Filament v3 and Filament Shield for complex RBAC. Designed a comprehensive audit trailing system using Spatie Activitylog and developed advanced reporting capabilities via OpenSpout. Enforcing absolute security across sensitive endpoints, designing dynamic summary rows for complex business intelligence XLSX exports, and implementing secure user impersonation workflows for seamless administrative support without compromising audit logs. Enhanced overall platform security by integrating Google 2-Factor Authentication (2FA) and detailed device tracking. Established an automated testing suite with Pest and drafted extensive technical documentation, ensuring high software quality and seamless client handovers.

Click card to read full summary

Summary

The organization required a unified platform to consolidate disjointed administrative workflows. Managing student enrollments, tracking teacher schedules and attendance, processing payroll, and monitoring financial health required an integrated, automated system with role-based access. Architected a full-scale ERP using Laravel 11 and Filament v3. Designed highly relational database schemas covering five main domains: Academics, HR/Payroll, CRM, Finance, and Project Management. Implemented dynamic administrative interfaces with advanced data tables, calendars, and date range filters for reporting. Modeling complex data relationships between class schedules, teacher hourly rates, and automated payroll calculations. Implementing robust role-based access control (RBAC) to ensure that staff, teachers, and administrators only accessed their specific operational domains. Delivered a centralized management hub that significantly reduced administrative overhead. Streamlined student lifecycle management from enquiry to enrollment, automated attendance-to-payroll workflows, and provided comprehensive financial dashboards for leadership.

Click card to read full summary

Summary

Maintaining and scaling a complex multi-vendor e-learning platform required continuous improvements to the vendor and administrative dashboards, as well as optimizing the underlying API business logic for complex financial transactions. Focused heavily on backend API maintenance and full-stack development of the Admin and Vendor panels using Laravel 9 and PHP 8.1. Improved and optimized complex business logic governing installment payments, dynamic discount rules, affiliate tracking, and the multi-vendor accounting module. Ensuring absolute mathematical precision and strict transactional integrity when processing cart-level discounts, cashback rewards, and dynamic vendor commission payouts across the platform. Successfully stabilized the platform's API endpoints, delivered streamlined administrative and vendor dashboards, and ensured highly accurate and performant financial business logic execution for thousands of daily transactions.

Click card to read full summary

Summary

Standard vector searches produce generic responses. The goal was to build an intelligent, bilingual travel assistant capable of contextual, intent-aware recommendations, deciding dynamically when to fetch live data via external tools versus querying internal knowledge bases. Engineered a fast Python/FastAPI backend executing an advanced pipeline (language detection, query rewriting, intent routing). Integrated a hybrid search system combining dense vector (Qdrant) and sparse keyword (OpenSearch) retrieval, refined by an ONNX INT8 optimized reranker. Built a highly interactive Next.js streaming frontend via Server-Sent Events (SSE). Orchestrating real-time execution of external travel tools, sanitizing massive JSON outputs, and performing live currency conversion to MMK. Synchronizing streamed LLM text responses (Gemini 2.5) with delayed rendering of structured interactive 'Place Cards'. Delivered a robust, highly accurate AI travel assistant with intelligent fallbacks and multi-turn conversational memory using Redis. Successfully containerized the entire ecosystem (API, VectorDB, SearchDB, Cache) using Docker Compose for seamless deployment.

Click card to read full summary