PROVEN PARTS → ASSEMBLED, NOT REBUILT
Production-deployed AI systems, developer frameworks, and security research — built with one engineering practice. Each entry below states its status explicitly: running in production, published, or research.
PROVEN PARTS → ASSEMBLED, NOT REBUILT
The hard parts — auth, AI, payments-grade data, delivery, marketing — are already built and tested once, as shared building blocks. A new product doesn't start from zero: it starts most of the way done.
A natural-language specification becomes a deployed app through a six-gate pipeline (ADL), with a five-model consensus engine cross-checking every AI decision.
A unified AI client, shared authentication, config, audit, email and a common UI kit — each one production-tested in live apps and reused by the next.
One toolchain builds each app, generates its store screenshots, pushes to TestFlight, and releases to the App Store and Google Play.
The same studio generates narrated, captioned promo videos and audio explainers from a storyboard, with AI B-roll on demand.
Every checked block is already built and tested in production. A new product assembles them instead of rebuilding — so it starts most of the way done and ships in a fraction of the time.
SpendCity
LIITOS
Topos
BoxSight LLC is a USA-based technology architecture firm. Nine capabilities, one engineering practice.
SpendCity
Live deployment
Reads documents entirely on your phone: on-device AI, no server, no account. Your data stays encrypted on the device. Features a gamified "Financial City" visualization.
LIITOS
Live deployment
News and research intelligence in one app, kept as separate sections: daily AI briefings built from curated news sources, with AI audio episodes several times a week, a weekly deep dive and a monthly video briefing, beside a sustainability research section built on university-affiliated feeds, with a weekly podcast and a monthly video briefing.
Topos
Live deployment
Tap-first nearby-place finder where travel time replaces distance as the search dimension. Two taps — category plus time radius — return matching places with AI-generated result labels, 22 pre-built situation bundles, and multi-stop trip chaining. Live on the App Store and Google Play.
READ THE SPEC → PROVE THE DIVERGENCE
Two-primitive specification language (Entity + Transition) with a 6-gate LLM pipeline and 5-model consensus mechanism. Moves from natural-language intent to deployed application across six target stacks. Used to build six applications with zero human coding after intent submission. ADL-UI — a browser-based spec session that generates a preliminary build plan — is coming to adl.boxsight.ai.
Read the ADL whitepaper →Multi-model consensus engine using the Model Context Protocol. Queries 5 LLMs simultaneously with Jaccard similarity clustering to surface disagreement and reduce single-model failure modes. Published on GitHub and npm.
A methodology for auditing the rules that hold the internet together — IETF RFCs, W3C specs, OAuth/WebAuthn/JOSE flows — by reading what they MUST/SHOULD/MAY require and then proving where real-world implementations diverge. HUNTER's first novel CVE shipped in 2026: CVE-2026-48522 — a previously-unreported divergence across the JOSE ecosystem (Python/Java/Node).
Now spanning eleven phase milestones across JWT/JOSE, X.509, OAuth 2.0, WebAuthn, CBOR/COSE, DANE (TLSA), TLS 1.3 trust-anchor handling, SAML, and others — each run through five gates before any divergence claim is made.
Read the HUNTER capability brief →Local-first network security monitoring. Distributed agents running local LLMs (Ollama) produce plain-language threat detection and network health reports. No telemetry leaves the perimeter.
Pre-flashed USB for zero-configuration network OS deployment. Plug in, pick an OS, walk away. Installs a PXE boot control center alongside any existing OS with cloud licensing, auto-updates, and telemetry. Supports Linux, macOS, and Windows targets via iPXE, GRUB2, and WinPE.
This is deliberate architecture, not accidental reuse. Every capability built for one product — vector search, LLM reasoning, multi-model consensus, on-device inference — deploys to the next without re-engineering, so each app ships faster than the last. The same spec-driven methodology even extends to bare-metal infrastructure in LINETBOOT-USB. The studio is the asset; the apps are what it produces.