Hire me: local, private AI builds

I'm Benjamin Hayward (AI Playground LLC). I design and build local, offline AI tools for people who want to own their stack: they run on your hardware or in your own cloud account, with no platform cut and none of your data stored with me.

Shipped: VIP Chat, Diagram Tool, DJ Radio, Music Master, MTA music and AI films, Power OCR (live demos by appointment) and APAM (paper trading only; not financial advice) and StoryForge. In progress: JamLocker is coming soon to the App Store. Below is what I have built, shipped or tested, and the skills each one used. Every job starts with a free consultation.

Recognition

Accepted into NVIDIA Inception (NVIDIA's program for AI startups), 2026. Eight projects shipped: VIP Chat, Diagram Tool, DJ Radio, Music Master, MTA music and AI films, Power OCR, APAM and StoryForge. Next: JamLocker, coming soon to the App Store.

AI work

Power OCR

Live · setup after a free consultation

A private document reader for offices such as law firms. It reads scans and PDFs, indexes them, and answers questions only from the documents you pick, citing the file and page. It runs in your office or in your own cloud account, so your files stay with you.

  • Built on a fork of an open-source OCR model. I audited it first, removed the hidden call-home code, and kept the original license and credits.
  • Tested in a sealed Docker run with no network on one RTX 5090: 782 real scanned court pages at 78.4 pages a minute as one document, or 233.5 a minute with 8 running at once, with no errors.
  • Known limit from that test: very small print on product-label exhibits was missed. Ordinary filings read well.

Skills used: OCR models, SGLang serving, Docker, GPU tuning, security review, document Q&A with citations

Power OCR page: the app, example build and time estimator

Local models and VIRC

Ongoing

I run AI on my own hardware where I can.

  • Local models through LM Studio and ComfyUI, and Nemotron and Gemma as stand-in reasoning cores in VIRC.
  • VIRC: a reasoning runtime that may only use mounted, signed knowledge packs. Its Stage 1 test set ran with 0 of 39 attack cases succeeding.

Skills used: LM Studio, ComfyUI, local LLMs, Ed25519 signing, audit logging, threat testing

StoryForge

Shipped

An AI production pipeline for books and audiobooks: from series plan to prose, then images and voiced audio, with review gates at each stage.

  • Produced the full-cast audio of LitRPG Book 1, all 12 chapters, through the pipeline.
  • Stages can be checked, replayed and rerun on their own without breaking the rest.
  • Adapted Book 1 into a TV script and a movie script, and made a first film test of one scene on a local GPU.

Skills used: LLM pipelines, ComfyUI, text-to-speech and voice cloning, Supabase, human review gates

StoryForge page: each step with real samples from Book 1

MTA music and AI films

Live

Marching Tavern Anthem (MTA) is a music style I created, mixing military cadence with tavern folk singalongs. It's also my marketing workshop. I make the songs, films, Shorts, thumbnails and promo campaigns myself by combining image, video, music and voice AI tools in ways no single tool can do alone. I can build marketing videos, product films, Shorts, music and promo assets for you, localized into many languages, from one idea to a finished campaign.

  • Released songs on Spotify, Apple Music and YouTube Music, with music videos and short films on the YouTube channel.
  • Launched films in 15 languages beyond English, each with its own premiere schedule, its own Short and localized titles and metadata.
  • Built a full pipeline: first- and last-frame video in ComfyUI, local music generation with MiniMax-Music3 and ACE-Step, ElevenLabs voices, 4K upscaling and mastering.
  • Built the V-MTA training workbench: I reviewed 2,000 candidate clips, kept 193 and ran LoRA fine-tuning experiments.

Skills used: AI video, image and music generation, multilingual localization, YouTube growth and analytics, dataset curation, LoRA fine-tuning, audio mastering

MTA music and AI films on this site, or open Marching Tavern Anthem on YouTube.

Hire me for it: Marketing videos, Shorts and promo films

APAM (Adaptive Price Action Model)

Shipped · paper trading only

A local paper-trading research tool that runs on your own PC. A fleet of trading bots trades on paper using market data from a key you provide. The best are promoted and the weakest are retired. No cloud account, and none of your trading data is uploaded.

  • Live-feed side shipped: dashboard, watchlist, charts, AI recommendations (a local LLM you run, such as through Ollama) and live signals.
  • Historical replay on a unified dataset covering a market index back to 1871 and daily stocks from 1962, with blackout-block gap handling and survivorship-bias warnings on pre-1996 data. Replay is built to avoid lookahead bias.
  • Human oversight controls and local-first data in SQLite. It places no real trades and has no live order execution. Not financial advice.

Skills used: Autonomous agents, local LLMs, market-data APIs, simulation, evaluation design, SQLite

APAM page: live feed, historical replay and how it stays on your PC

Security

AI tool security reviews

Ongoing practice

Before I run a downloaded AI model or tool, I read its code for anything that phones home, runs remote code or leaks data.

  • Power OCR fork: audited the upstream OCR code and its dependencies, then removed or switched off remote-code auto-updates, library telemetry, eval() on model output, unsafe pickle loading, and page snapshots and document text written to disk and logs. Each change is documented with where the data went.
  • Power OCR runs offline: model files are checked against SHA-256 hashes before loading, a network guard blocks outside connections, and speed tests ran in Docker with networking turned off.
  • New tools such as Mirage are staged in isolation with verified file hashes and a static review before anything runs. That review found Mirage's Python SDK sends telemetry by default, so it is kept with telemetry turned off.

Skills used: Static code review, dependency audit, egress testing, Docker network isolation, SHA-256 verification

Secure-by-default products

Ongoing practice

My products collect as little as they can, and keys stay out of code and chat.

  • Creator licenses are Ed25519-signed, and the license service only accepts Stripe webhooks with a valid signature.
  • Secrets live in environment variables or Cloudflare Worker secrets and are never printed or committed. Diagram Tool checks subscriptions with a read-only restricted Stripe key.
  • No-data design: Diagram Tool keeps files in the browser, JamLocker has no login and no tracking, and the chat stopped storing payment history, Stripe IDs and IP addresses.
  • Chat database locked down with row-level security and tightened table and function grants.
  • File moves are SHA-256 verified before and after, with a log that can reverse any move.

Skills used: Ed25519 signing, Stripe webhook verification, least-privilege API keys, secrets management, Supabase RLS, privacy-by-design

Payments and infrastructure

VIP Chat and creator licensing

Live

The live chat on this site, packaged so other streamers can host it themselves. General chat is free; $5/month unlocks VIP so you can earn, and you keep all of it (no cut).

  • Stripe subscription checkout, signed webhooks, and a customer portal that cancels at the end of the paid month. Tested end to end with a live payment on Nov 11, 2026.
  • License service on Cloudflare Workers and KV: Ed25519-signed keys and 5-day leases that each install checks daily. If a license lapses, only VIP locks.
  • Runs on the creator's own PC with Docker, or their own cloud account. I store none of their chat data.

Skills used: Stripe, Cloudflare Workers/KV, Supabase, Docker, React, TypeScript

VIP Chat

Diagram Tool

Live

A diagram editor that opens and exports Word, PDF, draw.io, SVG, Visio and image files. Files stay in your browser. First hour free, full editor with no limits, then $5/month.

  • Live with Google and Apple sign-in and a $5/month Stripe subscription. The first hour is the full editor, free, with no card.
  • Flowchart, network and cloud icon sets (AWS, Azure, Google Cloud), with hosted smoke tests run after each deploy.

Skills used: JavaScript, Cloudflare, OAuth sign-in, Stripe, file-format conversion

Diagram Tool on this site, or open Diagram Tool on its own site.

JamLocker

Coming soon to the App Store

An offline iOS music app for songs you own. No login, no tracking and no subscription. You buy consumable song packs.

  • Shipped EAS builds to TestFlight with StoreKit consumable song-pack purchases.

Skills used: React Native, Expo, TypeScript, StoreKit

DJ Radio

Live

A standalone radio app that runs on your own PC. It plays from your music folder, puts a player page in OBS, and lets fans vote on the next hour's genre over a link. A $5/month key runs the whole station; free 12-hour demo (enough for one YouTube stream).

  • Standalone v2 is finished: dashboard, seek slider, library and Up Next, OBS player, fan votes and requests, and a clear way to start and stop.

Skills used: JavaScript, OBS browser sources, audience voting

DJ Radio

MTA Music Master

Live

A local mastering tool that runs on your own PC. Point it at a folder of songs, or drag them in, and it masters each track to a genre preset. Nothing is uploaded. $5/month, or $9/month with DJ Radio on one key.

  • Sold through the same Stripe checkout and license service as DJ Radio and VIP Chat: a $5/month key, or a $9 combo key that also unlocks DJ Radio, with a prorated upgrade from either single plan.

Skills used: Stripe, Cloudflare Workers/KV, Ed25519 licensing, local audio processing

Music Master

What custom solution do you need?

Small office or enterprise, it starts with a free consultation. Then you choose: buy a one-time setup outright and own it, or pay a lower setup rate plus a monthly retainer that keeps me on call for AI/ML tech support for our products.

Step 1

Free consultation

Free

Tell me what you need. My first question is whether you have your own AI hardware or want it hosted in your own cloud account.

  • Your goal, your data, and any privacy rules it falls under
  • Which of our products fits (Power OCR, VIP Chat, Diagram Tool, DJ Radio, MTA, APAM) or what a custom build would take
  • A written recommendation: one-time setup or support retainer

Option A

One-time setup

Quoted after your consultation

Buy the build outright and own it. I install and configure it on your hardware or in your own cloud account, and train your team.

  • Install and configure on your hardware or in your own cloud account
  • Tested on your own data before handover
  • Training and a written runbook

Option B

Support retainer

Lower setup rate plus a monthly retainer, quoted after your consultation

Pay a lower setup fee, plus a recurring retainer that keeps me on call for AI/ML tech support for our products.

  • On call for fixes and remote help
  • Model, engine and security updates, tested before they reach you
  • Tuning and scaling help as your use grows

Power OCR installs include ClamAV (GPL-2.0, run unmodified as a separate component) and use NVIDIA CUDA libraries under NVIDIA’s own licence terms. We build them on your machine, and you accept those third-party terms.

Book a free consultation

Prices depend on the build, your data and where it runs, so I quote them after the consultation.