Hello, world
I'm Vivian Demello, welcome to my blog.
Vivian Demello
Software Engineer
Software Engineer and Linux nerd building AI systems
About
Hey, I'm Vivian! You might know me as vynride. I got into programming when I was 13 years old and ever since then I've been fascinated by computers and how they work. Out of this curiosity I explored Linux systems, virtual machines, self-hosting.
For the past two years, I've worked professionally as an AI Engineer at various startups. I love building and owning systems end to end.
Outside of work, I enjoy competing in hackathons, competitive programming and tinkering.
Tech stack
Python
TypeScript
Go
C++
Bash
Next.js
FastAPI
TailwindCSS
AWS
Docker
GitHub Actions
Nginx
Git
Linux
Neovim
tmux
Experience
Software Engineering Intern
Reduced LLM input costs by ~90% with context caching and shipped an A/B experiment that lifted conversions by 8%. Also migrated the entire infra from DigitalOcean to AWS.
- Cut LLM input cost ~10× on the highest-volume chat path by implementing Gemini explicit context caching with a static-core / dynamic-tail prompt split, validated by a live API spike showing 99.8% of the prompt served from cache
- Classified all 163K users by life-topic with an embedding-based classifier, sharding a ~13-hour job to ~1 hour via crc32 partitioning; built a Metabase cohort dashboard that surfaced a ~50× conversion gap between engaged and non-engaged users
- Shipped a version-gated home-dashboard A/B experiment (FastAPI + Flutter) to ~2,600 users, lifting trial-to-activation conversion by +8%
- Built Android app features on Firebase (Firestore, Node.js Cloud Functions) and delivered FCM push notifications to ~12.8K Android users
- Migrated the platform end-to-end from DigitalOcean to AWS (ECS, EC2, S3, CloudFront), moving ~200 GB of data and databases with a ~5-minute cutover and cutting monthly infra spend ~20%; migrated CI/CD from GitHub Actions to Jenkins
DevOps & Backend Intern
Deployed 15+ AI and full-stack microservices on Azure, then led a zero-downtime move to a self-hosted VPS that cut infra cost ~15%. Set up Prometheus + Grafana observability, keeping availability above 99%.
- Containerized and deployed 15+ AI / full-stack microservices on Azure Container Apps with Docker and Kubernetes manifests, configuring autoscaling rules, health probes, and zero-downtime rollouts
- Led a zero-downtime migration from Azure managed services to a self-hosted VPS using Docker Compose and Nginx reverse proxy with TLS termination; cut monthly infrastructure cost ~15% while preserving production traffic
- Instrumented system-wide observability with Prometheus + Grafana (SLO-based alerting for latency, error rates, and resource use), and triaged memory leaks in long-running LLM inference workers, sustaining >99% availability
- Refactored GitHub Actions CI/CD pipelines with Docker layer caching and parallel jobs, reducing average build time by 15% and lowering deployment failure rate
Software Engineering Intern
Built a RAG pipeline over 1.2K+ documents that lifted answer relevance 38% and top-3 precision from 0.61 to 0.84. Streamed LLM output cutting time-to-first-token (TTFT) 40%.
- Architected a production Retrieval-Augmented Generation pipeline using LangChain and ChromaDB, indexing 1.2K+ documents with hybrid semantic + keyword retrieval; lifted answer relevance 38% on a held-out internal eval set
- Designed a layered retrieval strategy combining recursive chunking with overlap-aware boundaries, cross-encoder reranking on top-K candidates, and HyDE-style query rewriting; raised top-3 retrieval precision from 0.61 to 0.84 and reduced grounded-query hallucinations
- Implemented token-level streaming for LLM outputs over Server-Sent Events with backpressure handling; cut Time-to-First-Token (TTFT) by 40%
- Built async FastAPI backend with pagination, rate limiting, and request-level tracing for token usage and latency attribution per pipeline stage
- Built the Next.js + TypeScript frontend with SSR, ISR, and CDN caching (Lighthouse 98); authored PyTest unit + integration tests and GitHub Actions CI to 82% coverage
Full Stack Developer (Contract)
Designed and developed the Cards20 landing page for an autonomous field-sales agent platform, with an interactive simulation of specialized subagents.
- Designed and built the Cards20 landing page, a command-center-styled marketing site for an autonomous field-sales agent platform
- Built an interactive simulation where users deploy and configure the platform's agent units (routing, calling, pitching, and objective analysis) over a sample field-sales mission
- Shipped the responsive Next.js frontend on Vercel with an animated command-center UI and per-unit context persisted across the simulation
Projects
Redline
A real-time voice roleplay trainer for crisis communication, with ~1.5 s voice round-trips over a single WebSocket. Personas branch across 20+ scenarios, and an LLM coach scores each transcript.
- Built a real-time voice-roleplay incident trainer, engineering a turn-based WebSocket loop that streams mic audio through Saaras STT → Sarvam LLM → Bulbul TTS with live severity and escalation state on the same socket, holding ~1.5 s end-to-end voice round-trips (mic to spoken reply) and ~600 ms time-to-first-audio via streaming
- Drove branching persona behavior via JSON-schema structured LLM outputs across 20+ scenarios, with an LLM coaching pass scoring each transcript on five dimensions; shipped on PostgreSQL + Docker Compose + nginx with OAuth
- Next.js
- FastAPI
WebSockets
PostgreSQL- Docker
RecallAI
Turns exam PDFs into analyzed reports through an async OCR + Gemini pipeline with live progress updates. Reached 100+ monthly users and 600+ report exports within 4 days of launch.
- Built a full-stack multimodal AI platform that ingests exam PDFs through an async FastAPI + Celery pipeline (PyMuPDF → Tesseract / EasyOCR fallback → Gemini → WeasyPrint) and streams live job status to the browser over Redis-backed Server-Sent Events; a 25-minute watchdog auto-fails hung jobs
- Hardened the BYOK model with per-user AES-encrypted Gemini keys behind an HS256 JWT proxy and Auth.js v5 OAuth; shipped via Docker Compose with Jenkins CI/CD, reaching 100+ monthly users, 1K+ page visits, and 600+ report exports within 4 days of launch
- FastAPI
Celery- Next.js
Gemini- PostgreSQL
Redis- Docker
Newstinct
Ingests and deduplicates ~25K news articles and serves 1K+ daily requests. A Redis caching layer cut P95 query latency by ~30%.
- Built a live product that processes ~25K articles and serves ~1K+ daily requests, with a horizontally scalable ingestion pipeline using Selenium and BeautifulSoup and content-hash + cosine-similarity deduplication against a PostgreSQL-backed embedding store
- Persisted ranked output with idempotent, retry-safe writes and added a Redis caching layer that cut P95 query latency by ~30%; wired end-to-end CI/CD via GitHub Actions with preview deployments
- FastAPI
- Next.js
- PostgreSQL
- Redis
Selenium- BeautifulSoup
NanoCraft
Turns any DIY guide URL into a step-by-step visual walkthrough with generated images for each step. A chatbot grounded in the guide answers questions about materials and steps.
- Turns any DIY guide URL into a structured step-by-step visual walkthrough via Gemini structured-output prompts
- Generates per-step scene images with Cloudflare Workers AI through async background jobs persisted to MongoDB GridFS
- Context-aware chatbot grounded in scraped project data answers material and step questions
React- FastAPI
- Gemini
Cloudflare Workers AI
MongoDB
Achievements
- Winner at JPMorgan Chase Code for Good '26
- MLH Vultr Track Winner at Hackprix '26
- 3rd prize at Technex 2025 among 400+ teams
- 5th prize at SITnovate 2025 among 200+ teams
- Special Mention at Murf AI May Buildathon
- Finalist at Paytm AI Hackathon
- Semi-finalist among 3000+ participants at the 100xEngineers Generative AI Buildathon
- Finalist at HackByte 3.0, IIITDM Jabalpur
- 2★ on CodeChef, max rating 1550
- Rank 3 at Geek ICPC 2.0, RBU
- 300+ DSA problems solved across competitive programming platforms
- Technical Co-Lead at The CodeBreakers Club RCOEM, mentored juniors on DSA, system design, and AI/ML fundamentals