Every January I make the same kind of New Year's resolution: cut down what I pay for the services I use sporadically, and clean up the sprawl of accounts, files, and tools I never quite get around to organizing. The list grows every year. This site is where I track what I've actually built, what's in flight, and what I'm still scoping out.
I'm not against SaaS — I happily pay for tools I use every day. The problem is the “higher tier” features I only need a few times a month, where the cheapest plan that unlocks them costs more than building it myself. So I build — and the AI assistants that used to be priced per-subscription are now available pay-per-use through AWS Bedrock, which makes a lot of this practical for the first time. I switch between Claude, the GPT family, and whatever else is best-in-class today without committing to any one provider.
Most of what's here lives under the tokenburner-ai GitHub org. Everything is public, MIT licensed, and designed to install into your own AWS account in dev mode for around a dollar a month idle.
The umbrella for everything else. Clone the stack repo, point an AI coding assistant at it, and ten minutes later you have a dashboard URL in your own AWS account with one card per installed feature.
S3-backed personal file storage with a CloudFront UI. Replaces my Dropbox subscription — same workflow, lives in my own account, costs a few cents a month.
Private AI chat that streams from Bedrock and stores history in DynamoDB. No third party reads my prompts, no big company grifts the conversations for training data.
Threaded discussion board with threads stored as JSON in S3, so an AI assistant with read access can use the forum directly as project context.
Account console and shared-context store. Provisions IAM users with scoped Bedrock access and serves an MCP endpoint, so my kids' Macs run agentic tools with guardrails and only the models I've approved.
A Shopify-ish POS with the financial pieces wired together end-to-end — receipts, taxes, employee discount tiers, receiving against purchase orders, inventory adjustments, customer-facing display. Built for a real cigar lounge that I help operate.
A free iPhone speedometer app, plus a small service that pulls from Apple Health and the Watch and renders the kind of activity views I'm currently paying Strava for.
A cloud-accessible LLM router that runs on hardware at home, so I can dig into the actual problems of model management, routing, and GPU utilisation rather than just reading about them.
Applying agentic systems to ASIC RTL verification — specifically formal assertion authoring and refinement, where LLMs can iterate against the formal solver in a loop instead of one-shotting an answer.
A personal trading and IRA-management application built on top of Interactive Brokers' API. Position tracking, rule-based rebalancing, and the analyses I keep wishing my brokerage site exposed.