Projects

End-to-end case studies: problem, product experience, and technical depth.

GenAI-powered Smart Maintenance Assistant
RAG chatbot for real-time troubleshooting
User problem
Maintenance teams were spending too much time hunting through logs, manuals, and sensor data to diagnose issues. Context was scattered across systems, and answers were slow—especially during incidents when speed mattered most.
UX / UI approach
A familiar chat interface so users could ask in plain language. Responses surfaced concise summaries plus source references (log snippets, manual sections, relevant sensor readings) so technicians could trust and verify answers. Streamed replies to reduce perceived latency and kept conversation history for follow-up questions.
Frontend architecture
React SPA with a message list and an input tied to a streaming API. State for conversation turns and loading/error UX; optimistic updates for sent messages. Responsive layout so it worked on tablets in the field. Accessible form and live region for streamed content.
Backend / system design
RAG pipeline: ingestion of maintenance logs, manuals, and sensor data into a vector store; retrieval of top-k chunks by query embedding; LLM (e.g. GPT/Claude) with retrieved context to generate answers. APIs for chat completion and streaming, with auth and rate limits. Chunking and embedding strategy tuned for technical content.
Impact and metrics
Faster time-to-answer for common troubleshooting queries; fewer context switches between tools; positive feedback from early users on relevance of retrieved sources. Qualitative wins on onboarding—new technicians could self-serve from documentation and past resolutions.
Learnings
Chunk size and overlap heavily affect retrieval quality for technical docs. Balancing latency (embedding + LLM calls) with completeness led to clear streaming UX. Tracing and evaluating LLM outputs against known good answers helped iterate on prompts and retrieval.
Real-time payment alerting platform
High-throughput APIs and observability
User problem
Operations and client-facing teams needed timely, reliable alerts on payment and transaction states. Delays or gaps in alerting led to reactive support and manual checks. Data lived in high-volume transactional and XML message flows that had to be parsed, validated, and surfaced without adding latency.
UX / UI approach
Clear alert lists and filters (by status, client, time) and minimal steps to acknowledge or escalate. Status and summary at a glance; drill-down into underlying messages when needed. Designed for both real-time monitoring and after-the-fact review, so the same UI served on-call and compliance use cases.
Frontend architecture
React app with tables and filters backed by REST APIs. Client-side state for filters and pagination; optimistic updates where appropriate. Careful loading and error states so users could tell whether data was stale or the system was degraded. Built to work within internal design system and accessibility standards.
Backend / system design
Java services for ingestion, parsing, and validation of financial messages; REST API layer for querying alerts and metadata. Emphasis on throughput, idempotency, and clear error handling. Observability (logs, metrics, distributed traces, Prometheus) for debugging and capacity planning. Backward-compatible releases and collaboration with DevOps/SRE for rollout and rollback.
Impact and metrics
Roughly 30% improvement in API response times and system stability after scaling and optimization. Lower MTTR during incidents thanks to better traces and alerting. Fewer support escalations from timely, accurate alerts.
Learnings
Structured logging and correlation IDs paid off when debugging cross-service flows. Investing in parsing and validation upfront reduced invalid states downstream. Balancing backward compatibility with iteration required clear contracts and incremental rollout.
Distributed approval workflow service
Event-driven orchestration and reliability
User problem
Approval flows were manual, multi-step, and brittle—hard to audit and easy to block on a single person or system. Operations wanted a consistent, traceable path from request to decision, with clear rollback and recovery when steps failed or timed out.
UX / UI approach
Workflow status visible end-to-end: pending steps, who (or which system) is responsible, and history of decisions. Notifications and queues aligned to those steps so approvers had one place to act. UX focused on reducing noise and making the “next action” obvious, while still supporting auditors who needed full history.
Frontend architecture
React views for workflow dashboards and approval queues, wired to REST and optionally to real-time updates. Role-based visibility so users saw only relevant queues. Built with the same design system and accessibility patterns as the rest of the stack. Performance considerations for large result sets and long-running workflows.
Backend / system design
Event-driven service with clear state machine for approval steps. Durable execution, retries, timeouts, and failover so in-flight workflows could recover from failures. Database transactions and idempotency keys to avoid duplicate actions. Refactor from monolith to services improved isolation and made it easier to add new steps and integrations.
Impact and metrics
Significant reduction in manual operations (on the order of ~40%) by automating routing and reminders. Support incidents related to approval flows dropped sharply (~70%) after hardening and better error handling. Auditors could rely on a single trail for compliance.
Learnings
Modeling workflows as explicit state machines made edge cases (timeouts, retries, manual overrides) easier to reason about. Decoupling services reduced blast radius but required clear event contracts and schema evolution. Investing in operational visibility (logs, metrics, traces) was essential for onboarding and incident response.

Certifications

Certified AWS Cloud PractitionerEarned January 2025

GenAI-powered Smart Maintenance Assistant

Real-time payment alerting platform

Distributed approval workflow service

Certifications