How We Build.

Layer on, don’t rip out.

We integrate with what you already run. Cameras, intercoms, venues, payment processors, identity providers, NVRs, learning management systems — whatever is in place is in place. The new value is in the operations on top, not in replacing the layer underneath. When a customer has invested in hardware or process, the answer is “we extend that,” never “you switch.”

• QORAPlugs into existing ONVIF/RTSP cameras + Agent DVR + iSpy. We don’t sell cameras. The VMS keeps doing recording. We do triage on top.
• VenueFuzeOrganizer picks PayPal, Stripe, or Square — we don’t force a payment processor. Works with whatever venue you already booked.
• NGChessTeachers don’t leave their classroom management tool. NGChess hands progress data back via simple join codes and CSV exports.

The shortest way to say this: nobody’s going to rip out their cameras to use an AI tool. So we wrote the AI tool that doesn’t make them.

AI in the loop, humans in the chair.

Models propose. Humans approve consequential calls. AI is great at triage, summarization, drafting, dispatch routing, and explanation. AI is not great at deciding to put a police car in motion, accept a donation, fail a student, or settle an irreversible group decision. Every AI-driven action across our products is reviewable, reversible, and surfaces the model’s reasoning.

• QORAModel proposes “Dispatch.” Operator clicks. We do not auto-dispatch — that’s how police roll on a deer.
• VenueFuzeAI drafts the post-event email. Organizer edits. Organizer sends. Drafting is the job, sending is the human’s.
• NGChessThe AI tutor explains why a move was strong or weak. The student still plays. The teacher still grades. The tutor doesn’t.

When an AI action is automated (e.g. routing a low-priority incident to a different queue), it’s logged, attributed to the model + version, and one click away from being overridden.

Operations-grade, not demo-grade.

We ship to operators running real shifts. That means audit logs, role-based access, count parity across every page, observability you can put on a dashboard, and uptime that holds during World Cup–scale traffic. Demos can fake all of this. Operators can’t. We design assuming the latter.

• QORAEvery operator action and every AI decision lands in an immutable audit log keyed to user, model version, and timestamp.
• VenueFuzeOffline-first gate check-in. WiFi at the venue dies? Cached ticket list, local scan, reconcile when connectivity returns.
• QORA dashboardCustomer / site / camera / incident counts match across every page. New count tiles use server-side aggregates, never client-side filters on a paged list.

Buy-vs-build, honestly.

Build the layer where the value is. Buy or borrow everywhere else. Claude, GPT, open-weights models, hosted vector DBs, off-the-shelf MCP servers, Postgres, Stripe, Cloudflare, Let’s Encrypt — we use what works. Customers don’t care which model wrote the summary. They care that the summary is good, available, and affordable.

• QORAEdge inference on small open-weights models. Cloud summarization on hosted Claude/GPT. Vector search on Postgres + pgvector, not a separate cluster.
• VenueFuzeThree payment processors via official SDKs. Cloudflare for DDoS. PostgreSQL for everything else. We didn’t build a payment processor.
• Zafronix APIsOpenAPI spec, standard JWT auth, standard rate-limit headers. The novelty is the data, not the protocol.

Measure the operation, not the model.

An eval score is not a business outcome. The question isn’t whether the model scores 91% on some benchmark. The question is whether dispatch happens in 8 seconds instead of 80, whether the event check-in line moves from 11 minutes to 2, whether the chess student went from 1200 to 1500 ELO over a term. We instrument the operation, not the model, and that’s the number we show customers.

• QORAPilot site: capture–to–dispatch went from 14 minutes to under 90 seconds. Same humans. Same cameras. Different layer on top.
• VenueFuzeGate scan time: 30 sec/person clipboard → 1.5 sec/person QR. The wine stays cold.
• Zafronix APIsMedian request latency, p99 latency, uptime — surfaced on a public status page. No vanity metrics.

Honest about what we’re not.

Real-time is hard. Betting-grade SLAs are harder. Some things we won’t do, and we say so upfront. Zafronix APIs run with a deliberate 5-minute live-match lag — we’re not Sportradar and we’ll tell you so the first time you ask. QORA doesn’t prevent crimes; it surfaces incidents faster. VenueFuze isn’t a CRM. Saying what we’re not is cheaper than disappointing a customer who assumed otherwise.

• Zafronix APIs5-minute lag during live matches, documented prominently, with a dedicated blog post explaining why.
• QORAWe don’t pretend the model is right 100% of the time. Operator can dismiss any incident. Dismissed incidents train the next model.
• NGChessFree tier exists and stays free. We don’t pretend the premium tier “makes you a grandmaster.” It gives you more games and deeper analytics.

LLMs — the brain.

An LLM (Large Language Model) reads and writes language. Give it text in, get text out. It can summarize a 200-page document, draft an email, translate, classify, explain, or answer questions. Under the hood it learned patterns from a vast amount of training text, and on top of that foundation it can reason about, plan, and analyze problems expressed in words. Think of it as a very well–read intern: fast, fluent, and surprisingly capable at reasoning through problems that can be expressed in language.

What it can’t do, by itself: look anything up in your data, take actions, click buttons, or persist knowledge across sessions without an external memory system. Without help, it only knows what was in its training set up to a cutoff date.

• QORAThe one-sentence incident summary on the operator console (“Person in restricted zone after hours…”) is an LLM writing English from a structured detection record.
• VenueFuzeThe follow-up email the organizer reviews after the event — LLM drafts, human edits, human sends.
• NGChessThe plain-English explanation under each move (“Nf3 develops a piece and controls e5/d4…”) is an LLM translating chess-engine output into something a 4th grader can read.

RAG — the brain with a reference library.

RAG (Retrieval-Augmented Generation) gives the LLM a way to look things up before it answers. Instead of relying on what was in training data, it first searches a private knowledge base (your incident history, your event archive, your API docs, your camera metadata), pulls back the most relevant snippets, and hands them to the LLM along with the user’s question. The LLM then generates an answer grounded in the retrieved information and its own reasoning capabilities — bringing your facts into the response instead of relying on general internet knowledge.

If LLMs are the brain, RAG is the brain with a reference library bolted on.

• QORA“Has this person tried this gate before?” — RAG searches the site’s incident history and feeds matches into the LLM’s summary.
• VenueFuzeRecognizing “Sarah, our top donor” across years of events — RAG over the venue’s historical attendee records, not generic ML.
• Zafronix APIsThe AI-enriched storyline field on every match (“First teenage hat-trick in a WC semifinal”) is RAG over our own structured tournament archive plus human-reviewed facts.

Agents — the brain that can take actions.

An agent is an AI system that can use tools, make decisions, and execute multi-step workflows toward a goal. Instead of just producing text, it can call APIs — send an email, route a ticket, schedule a dispatch, query a database, post to Slack — and then read the result and decide what to do next. Modern agents often combine a planner, a memory system, a workflow engine, and one or more models working together; the LLM is one component, not the whole thing. Agents loop: think, act, observe, think again. Done right, they replace a chain of human clicks that used to take ten minutes with a single sentence: “handle this.”

Where we’re careful: agents can dispatch responders, initiate financial transactions, or communicate with large groups of people. We follow Principle 02 here religiously — agents propose; humans approve consequential calls; every action is logged and reversible.

• QORAThe dispatch routing agent: classifies severity, picks the right responder team, drafts the message, queues the dispatch. The operator approves the send.
• VenueFuzePost-event agent: composes the thank-you email, drafts the donor follow-up, summarizes the night for the board chair, generates the tax-receipt PDF. Organizer reviews; agent sends.
• NGChessThe tutor agent watches a game, notices a pattern (“you keep falling for the same tactic”), and queues a personalized puzzle for tomorrow. Teacher can override.

MCP — how the brain plugs into the world.

MCP (Model Context Protocol) is an emerging open standard for connecting AI systems to tools and data sources. Before MCP, every product wrote its own integration glue for every tool the AI needed to use. MCP is a common plug-and-socket: a tool exposes itself once via MCP, and any compliant AI client can use it with minimal custom integration work. Think USB for AI — a single shape that lets the brain talk to your camera system, your payment processor, your CRM, or your knowledge base without bespoke wiring each time.

For customers, MCP means “our AI can use your stack without us building a custom connector for each tool you have.” That’s what makes Principle 01 (“layer on, don’t rip out”) economically realistic.

• QORAWe expect camera, NVR, and intercom vendors to increasingly expose MCP-compatible interfaces over time. Where available, we consume those integrations instead of writing custom connectors.
• VenueFuzePayment processor MCP servers let the post-event reconciliation agent pull receipts directly — no custom Stripe-vs-PayPal-vs-Square code.
• Zafronix APIsComing soon: a Zafronix APIs MCP server so AI agents built by anyone can query the World Cup family directly, with auth and rate limiting handled at the protocol layer.

If you stop reading here, this is the most important sentence on this page: none of the four layers is “AI” in the marketing–poster sense. They’re building blocks. Real products are mostly old software with carefully–chosen pieces of these layers swapped in at the right spots. That’s the job.