I design and ship production-grade backends — from low-latency AI verification pipelines to fraud detection systems. Every system I build is meant to be used, measured, and improved.
Not prototypes. Systems that solve concrete problems, with architecture decisions to back them up.
Microservice-based verification system for AI-generated content — built to catch what LLMs get wrong.
AI systems hallucinate. Existing solutions check outputs superficially. This needed a multi-stage pipeline that actually retrieves evidence and compares it semantically before scoring.
Designed a retrieval → semantic comparison → scoring pipeline across isolated Docker containers. Each stage is stateless, independently scalable, and communicates through clean interfaces. Redis eliminates redundant computations on repeated queries.
Real-time fraud detection system for fake job postings — from 120ms to 70ms, with a feedback loop that keeps learning.
Fake job listings cause real harm. Platforms needed a backend that could flag fraud in milliseconds, not after review queues.
Built async preprocessing pipelines to decouple NLP inference from request handling. JWT-secured endpoints with role-based access. A MySQL-backed feedback loop lets flagged posts feed back into model retraining — the system gets sharper over time.
A communication confidence app built around a core truth — you can't improve what you can't measure.
Most people avoid public speaking because they never get structured, private feedback. Coaching is expensive. StageFear makes the loop accessible.
Designed around a tight record → analyze → track loop. Users record themselves, receive structured feedback, and track improvement over sessions. The backend stores progress timelines, not just snapshots — because growth takes time to see.
Volunteer-senior assistance platform — with trust built into the architecture, not just the design.
Connecting senior citizens to volunteers sounds simple. But without identity verification and role control, the system becomes unsafe. Trust is a backend problem.
Built 4-role RBAC with Spring Security + JWT: Senior, Volunteer, Admin, Verifier. Admins manually approve volunteers after ID verification. REST APIs handle scheduling, real-time status updates, and service coordination — with clean separation per role.
Python Developer Intern
Built an end-to-end NLP-based fraud detection system, covering data preprocessing, model evaluation, and backend integration for real-time use.
"I focus on building systems that are efficient, scalable, and actually used."
I think about backend systems the way an engineer should — in terms of constraints, tradeoffs, and real impact. Not lines of code, but what breaks under load, what costs you latency, and what makes a system trustworthy enough to depend on.
My work spans NLP-powered fraud detection, AI content verification, and volunteer coordination platforms. Each one taught me something different about designing for the edges — not just the happy path.