Real products, shipped end to end.

01 · Mobile · AI Agent · Monetized

Tupliq Agent — an AI agent you can actually install

The problem

Remote workers, freelancers, consultants and small businesses don't want another chat window — they want outcomes. Tupliq Agent takes a plain-language request, classifies it, picks a workflow and the best available AI model, executes it, and returns structured, actionable results you can copy, save and reuse. It was built as a standalone product for the RevenueCat Shipaton challenge, with monetization designed in from the first commit.

How it's built

• Mobile client — Expo 54 / React Native with expo-router, Supabase auth, React Query + Zustand, and a role-based onboarding flow.
• Backend — a production NestJS REST API (tupliq-backend) with Prisma + Postgres: /agent/run orchestration, run history, usage limits, saved outputs, and RevenueCat-backed entitlements, containerized with Docker.
• Provider chain with fallback — Gemini ▸ OpenAI ▸ Anthropic. AI keys never leave the backend; the app only ever talks to the NestJS API.
• Trust nothing client-side — usage limits are enforced server-side (402 → paywall), and subscription truth is written only by the RevenueCat webhook, never read from the client SDK.
• Structured outputs — every workflow validates model responses with Zod; malformed output triggers one bounded repair pass instead of surfacing garbage to the user.

Outcome

A complete, monetizable vertical slice: signup → onboarding → run a workflow → history → free-tier limit → paywall → Pro entitlement via RevenueCat (monthly and yearly products behind the tupliq_pro entitlement), with the Android app published on Google Play.

A YouTube walkthrough is in production and will be linked here.

02 · AI Evaluation · Research

MechEval — grading frontier AI on real engineering physics

The problem

Frontier AI models can rapidly generate plausible-looking engineering derivations. But in safety-critical physical systems — civil infrastructure, aerospace airframes, high-pressure piping, machinery — a calculation that looks convincing but is physically invalid can mean structural collapse. Prior work shows LLMs can generate simulation inputs (e.g. GMSH geometry); MechEval asks the complementary question: can these models be trusted to get the physics right?

How it's built

• Code-graded, not vibe-graded — every task is scored by deterministic graders against a reference solution methodology, with rigorous validation of the grading logic itself.
• Five professional tasks — spanning mechanical and structural engineering, with defined input/output schemas per task.
• Reproducible by design — Python 3.11+, exact reproduction commands published, and every figure quoted from the run's own results/raw/*.json. 33 tests passing.
• Honest reporting — the README includes an AI-assisted development & human verification disclosure, plus known limitations and runtime/inference cost analysis.

Outcome

A measurable, reproducible result set across a live roster of frontier models — including the specific failure modes graders recorded — built for the micro1 Eval Hackathon 2026. The 2:23 narrated walkthrough reads every number directly from the benchmark's raw output files.

03 · Agent Tooling · Hackathon

Tupliq Launchpad — RevenueCat metrics inside your AI agent

The problem

If you ship monetized apps, your revenue data lives in RevenueCat and your release notes live in git history — both behind context switches. Built during the Hermes Agent Plugin Hackathon (Nov 2026), Launchpad brings both into the agent you already talk to.

How it's built

• Four tools, one plugin — launchpad_overview (revenue + actives + churn for the last N days), launchpad_revenue (explicit date ranges, gross/net/proceeds), launchpad_chart (any named RevenueCat chart: MRR, ARR, trials, conversion, …), and launchpad_changelog (drafts Play Store release notes from git commits since the last tag — read-only).
• Zero dependencies — stdlib urllib only. No Python packages to install or break.
• Secrets done right — the API key is stored via Hermes' credential route and never touches config.yaml; a V2 RevenueCat key is scoped to charts_metrics read only.
• Never raises — all handlers return JSON; network errors surface as readable error payloads instead of stack traces.

Outcome

Prompts like "How much revenue did Tupliq Agent make in the last 30 days?" or "Draft release notes for the Play Store update" now work in one turn — the same plugin powers this site's owner's day-to-day shipping loop.

04 · Simulation · Desktop · Rust

Tupliq Multiphysics — a workbench for co-simulation

The problem

Virtual commissioning — validating a machine's control logic against its physics before anything is built — still means gluing together mesh generators, solvers and PLC tooling by hand. Tupliq Multiphysics is a desktop workbench that wraps that pipeline in one app: Geometry → Mesh → Solve → Results, as a real numbered sequence rather than decorative tabs.

How it's built

• Tauri, not Electron — a Rust core with a React + TypeScript + Tailwind v4 frontend; the JS ↔ Rust round trip is already proven end to end with a stubbed run_solver(stage) command.
• A solver sidecar — the planned seam plugs in @agent-musaj/gmsh-generator and @agent-musaj/elmer-generator packages via a Tauri external binary, streaming solver stdout to a live log panel instead of blocking on returns.
• 3D viewport — a three.js / react-three-fiber panel reserved for geometry and mesh output from the pipeline's first stages.
• FMI/FMU + virtual PLC — the roadmap's hardest and least-served part: co-simulation standards and a virtual PLC loop with the least prior art.
• Drafting-table design — deliberately blueprint-flavored: fine grid, numbered pipeline, and a footer styled as an engineering drawing title block (DWG NO. / STAGE / REV / SCALE) that displays live run state.

Status

Active development. The frontend builds and type-checks clean; the next milestone is confirming the full Tauri toolchain locally, then wiring the solver sidecar.