Michael F. Chaves

An OpenAI-compatible LLM routing platform: prompt compression, PII redaction, cheapest-capable-model routing, budget ceilings, and auditable savings proofs.

What it does: any app that talks to an AI model can point at the Router instead. In-path, it compresses the prompt, strips personal data before anything leaves the building, sends each request to the cheapest model that can actually do the job, and enforces per-key budgets and per-request cost ceilings. Every call produces a receipt, so the savings claim is a ledger, not a slogan — this is the platform behind the measured 35–45% token-cost reduction on the résumé. The interface demos are live (link below); the hosted API is mid-rebuild. Source stays private: it is core IP.

Proves: backend & API engineering (Python / FastAPI) · LLM cost engineering · security posture · documentation discipline

Live interface demo

A working marketplace and analysis platform for AI inference — an inference book, live token pricing, and a proving ledger for genuine routed receipts — with DAR, its market-intelligence component, powering a DeFi analysis terminal.

What it does: AIM treats AI inference like a traded commodity. It tracks live token pricing across providers, keeps an inference book, and maintains a proving ledger designed to record only genuine, verifiable routed receipts — no synthetic numbers. It is a real TypeScript full-stack application (server, migrations, UI), currently in a ground-up redo so the public face matches the engineering underneath. DAR is AIM's market-intelligence component: its FastAPI backend — with an LLM router, user authentication, Stripe payments, a Supabase data layer, and PDF report export — generates market-analysis reports combining technical indicators, on-chain data, and sentiment. AIM's own catalog describes DAR as "Grok-powered daily crypto market analysis — the visual language AIM inherits." Its backend is built; its front end is mid-migration from its prototype. A compute-unit (CU) monetary engine — the ledger layer for AIM's internal economy — is under construction alongside it.

Proves: full-stack TypeScript product engineering · FastAPI service architecture · platform & market design · third-party integrations (payments, data, LLM)

The front door of the office stack: an AI reception supervisor and service desk for property shops and short-term-rental hosts.

What it does: MarciAi answers the phone in a voice the client chooses, trained on the client's industry and restricted to words the client approved. It qualifies callers, books appointments, and triages guest problems at 2 a.m.; emergencies escalate to the owner with the facts already gathered, and the owner receives supervised summaries instead of raw chaos. Its own one-pager states the status plainly — now onboarding hosts and clients — and nothing here claims a shipped product or a client result.

Proves: voice-AI product design · service design for a real trade · honest go-to-market discipline

A mobile-first PWA dashboard built for a working claims manager — in daily use by a real user, with her data kept on her own phone by design.

What it does: Sparks runs a claims manager's workday: labeled reminders with quiet hours (22:00–06:00) so nothing pings overnight; an inbox, queue, and log that keep her desk in one place; and an agent-fill import format behind a redaction fence that automatically blanks claim numbers, policy numbers, phone numbers, and email addresses before anything is processed. Requirements came from watching the job, not imagining it. The app itself is live at the link below — try it.

Proves: end-to-end product delivery for a non-technical user · privacy engineering · PWA engineering

Live app

A text-message filter built on a ternary method: messages from numbers not in your contacts are judged one at a time as they arrive — scams, spam, and threats quietly diverted, but you always know someone reached out.

What it does: Peace of Mind SMS never reads your inbox wholesale and never touches messages from your contacts. A text from an unknown number is judged on its own, the moment it arrives, and given a +1 / 0 / −1 verdict. Scams, political spam, and threats are diverted from your day — but never silently; you always know a message arrived. The spec includes a mobile-messaging threat registry and a production-hardening pass. Status, in its own README's words: concept / spec phase — nothing is built, shipped, or live yet.

Proves: consumer-safety product design · privacy-first architecture (judge one message, never read the inbox)

A future-developer platform for young creators: instead of a fixed curriculum, each creator gets a persistent virtual world that grows from their projects — and the city becomes a living portfolio of what they've actually made.

What it does: in NexGen, buildings are projects and progress lights the windows. Progression follows demonstrated capability and curiosity, not grade-level ceilings, and a guide character (Zali) walks with the creator. Safety and privacy documentation is part of the design from day one — written before any product code, on purpose. Today the work is platform design, prototype / facilitator structure, and those safety docs.

Proves: learning-platform design · youth-safety thinking · portfolio-based education design

"It writes what you meant." A speaker-intent correction layer for speech-to-text — requirements written by a man who runs his working life on dictation.

What it does: Sub_Scribe treats personal vocabulary — names, places, coined terms — as first-class data, subtracts the suggestion-takeover and worst-case readings dictation engines produce, and lands the text the speaker actually said, so dictation serves the speaker instead of the statistically likeliest word. It is a brainstorm-seed / requirements stage project (its README says "no code yet") — the requirements document itself is the artifact, and it is the one the big platforms never wrote.

Proves: accessibility product insight from lived experience · requirements engineering