AI Product & Venture Lead Β· Startup Founder Β· Sydney, Australia π¦πΊ
I play at the intersection of frontier AI, product strategy, and human-centred design. My background spans HCI and design, two startup exits, and 12 countries β now focused on building AI systems that actually work for people.
- AI Product Strategy β Defining product vision and roadmap for AI-native systems; from concept to deployment
- Human-AI Interaction β Designing experiences where AI augments human decision-making (built on Stanford HCI principles)
- Agentic AI Systems β Exploring multi-agent architectures, LLM orchestration, and AI-driven gameplay
- Data Science & ML β Bridging product thinking with model evaluation, feature engineering, and AI pipeline design
- Startup Building β Two successful exits in video games (Shanghai & Berlin); currently building with AI agents as co-founders
vivid-clean β local document hygiene for AI-assisted work
open source Β· privacy
A local-first tool for removing deterministic provenance markers and common AI-writing tells from documents, images and text. Built for people who use AI as an accessibility or writing aid and want control over the metadata attached to their work.
- Processes files locally and creates a cleaned copy rather than overwriting the original
- Supports Word, PowerPoint, PDF, plain text, Markdown and common image formats
- Checks for Unicode markers, document properties, C2PA data and EXIF/XMP metadata
Bash Β· Python Β· local-first Β· document hygiene
A curated slice of what I'm building. Full case studies and writing at vnsavitri.github.io.
vai_sante_os β privacy-first multimodal memory
active Β· research
A framework for provenance-aware multimodal memory and orchestration in high-stakes AI workflows (health, legal, policy, safety).
- Provenance-aware retrieval returns content and chain of custody
- Human-in-the-loop review gates for sensitive decisions
- Treats time and evidence quality as first-class, not metadata
Python Β· Mermaid Β· evaluation harness
dam-butler-mcp β Breville's first MCP tool, in daily production use
production Β· enterprise
GTM teams across APAC, North America, and EMEA needed daily access to 235K+ brand assets in Brandfolder β but retrieving the right file meant knowing the exact folder taxonomy, which most non-technical users didn't. Built Breville's first MCP-based internal tool: a custom GPT connected to the Brandfolder API via an intent parser and clarification loop.
- Natural language query β intent parser β structured Brandfolder API call
- Clarification loop resolves ambiguous inputs before the API fires
- Prototyped Sep 2025; shipped to production, in daily workflows across APAC, North America, and EMEA β demo video
MCP Β· Brandfolder API Β· ChatGPT Enterprise Β· Vercel
vivid-alpaca β multi-agent trading with execution guardrails
active Β· safety
Paper-first multi-agent AI trading lab built on the AlpacaTradingAgent lineage. Execution-layer guardrails sit between agent recommendations and broker order submission.
- Configurable agent mindsets (capital preservation β paper-aggressive training)
- Goal-aware workflows with target return, time horizon, max drawdown
- Live trading gated behind manual approval, journaling, cooldown, replay
Python Β· Dash Β· Alpaca API Β· multi-agent
espresso-horoscope-mcp β local-first MCP, OpenAI hackathon
shipped Β· hackathon Β· β 3
Local MCP project that turns espresso shot metrics into a personalized cosmic reading via GPT-OSS through LM Studio. Built for the OpenAI Open Model Hackathon (Best Local Agent category).
- 100% offline β no cloud inference
- Structured sensor data β strict tool/prompt boundary β user artifact
- Six-week deadline, 3-min demo video shipped
Python Β· Next.js 15 Β· LM Studio Β· MCP
sourdough-intelligence β pre-LLM data science, live product
live Β· vividcrumb.netlify.app
Started in 2018 β before LLMs. Built a two-stage model to find the sourdough recipe with the highest first-time success rate: multiple linear regression across recipe variables + IBM Watson NLP sentiment analysis on 207 recipes and their YouTube comment threads.
- Top-3 shortlist generated β picked one β worked first try
- Now a live scheduling app: bake-time wizard, baker's percentage formula gen, temperature-aware bulk fermentation, printable plans
R Β· IBM Watson NLP Β· regression
Almost β the life you didn't quite live
shipped Β· product
Upload a LinkedIn PDF. Almost finds 3β5 real fork points in your career history. You pick one. It renders the alternate you as a LinkedIn Ghost, Wiki Stub, Museum Plaque, or Tarot Card.
- Built on Anthropic Claude API (
claude-sonnet-4) - Claude native document support β no PDF library plumbing
- Four hand-tuned output formats, RevenueCat-gated Pro tier
Next.js 14 Β· Anthropic API Β· Fraunces
I use AI where it makes sense, part of the fun is figuring out whether AI would be helpful or slowing me down; especially if it produces slop. Ain't nobody got time to babysit dumb AI agents!
-
Open-source and local AI β My preference is to always try to use open and local models, choosing for privacy, cost, capability and the job at hand... if possible. Whereas at work, I kinda stuck with closed models from OpenAI or Anthropic (unfortunately).
-
Agent loops and graph engineering - I build the harness around agent workflows, with clear state, scoped delegation, useful hand-offs and memory that does not turn into a junk drawer.
-
Human-AI product design β I try to turn messy real-world work into AI products people can understand, trust and use.
-
Evaluation, evidence and safety β I test what a system does, keep provenance where it matters, and put review gates around decisions with real consequences.
-
Product direction and delivery β Moves from problem framing and prototype to a working product, with enough technical depth to make good calls along the way.
- AI-guided financial literacy app for next-gen Indonesians (early beta)
- AI-native video game development with agentic workflows and multi-agent collaboration
- Prototyping with Hermes Agent and isolated sub-agents for parallel task execution
- Running hybrid model workflows across local models and OpenRouter-hosted models, including Qwen 3.x variants
- Building AI evaluation frameworks for product decision-making
- Learning deeper Python for AI/ML, LLM fine-tuning, and agent orchestration patterns
- π AI Product Management β Duke University (Pratt School of Engineering)
- π MBA β Steinbeis Hochschule, Berlin (Thesis: AI-Driven Application for Experience Design)
- π©π»βπ B.A. in Science, Technology, and Society, Stanford University, School of Humanities & Sciences (HCI focus)
- π Lived and worked in 12 countries across 4 continents
- π£οΈ English Β· French Β· Mandarin

