Explore showcase
RAG-Powered Portfolio
Full-Stack & Artist · June 2025 – Present
Project2 / 13
Overview
At a glance
Personal portfolio and chat surface where retrieval-augmented answers stay grounded in my resume and project corpus. The build couples streaming LLM output with cost controls, Redis vector search, and a handcrafted pixel UI layer.
Next.js 15GeminiUpstash RedisVercel AI SDK✚
View GitHubView GitHub
Features
- Chatbot grounded in a personal project/experience corpus.
- Config-driven content architecture for fast updates.
- Edge-streamed responses with resilient error and retry-aware UX.
- Portrait idle animation, glasses toggle, dialogue variants, icon buttons, layered parallax.
- Distinct art direction with handmade pixel assets and motion polish.
Implementation
- Implemented embedding generation and Redis retrieval pipeline with cosine similarity ranking.
- Added SHA-256 query-embedding cache (24h TTL) to lower repeated embedding cost and latency.
- Implemented corpus hashing to auto-trigger re-embedding when underlying content changes.
- Added request-size and message-count guards for usage control and stability.
- Implemented Redis IP-based rate limiting with frontend-readable error semantics.
- Built typed hooks for typing state, timeline orchestration, responsive/parallax control, and interaction feedback pulses.
- Optimized media assets (including WebP compression) to reduce animation-heavy performance bottlenecks.
- Carried full cross-functional pipeline from Aseprite/Figma visual ideation to production code integration.
- Experimented with RabbitMQ during this project era and reused related queueing ideas later, though free-tier limits constrained continued usage.
Challenges
- Early performance was rough due to heavy parallax + animated menu combinations.
- Ongoing tradeoff between visual richness and runtime stability/performance.
- Security and durability still require hardening for larger traffic bursts.
- Free-tier limits on supporting infrastructure constrained some experimentation depth.
- Significant QA/refinement still needed for production-grade reliability.
Reflection
- This project is super personal to me. Seeing Aseprite -> Figma -> code turn into a living site over nearly a year was insanely rewarding.
- I put a lot of effort into tiny details because I wanted it to feel fun, not sterile.
- Fun fact: I made my own pixel GitHub icon before realizing there was already an official one.
- AI helped me move faster on boilerplate/type/error scaffolding, but the site still needs more hardening in perf/security/durability.
- I’m lowkey hoping traffic doesn’t spike too hard right now because I might be cooked.
- Still, performance improved massively after optimization passes and asset compression—my 2017 MacBook can handle it now!