The problem we set ourselves
AI products usually fail in one of two ways: they hallucinate (confident nonsense, zero grounding) or they hide (no track record, no accountability). We wanted to prove an AI product can do the opposite of both: make verifiable claims from live data, publish predictions before the event, and keep the receipts forever.
The proving ground: Formula 1, a domain with brutal ground truth. Every two weeks reality grades your model in front of millions.
Confident nonsense, zero grounding.
No track record, no accountability.
Verifiable claims from live data, picks published first, graded forever.
The architecture: deliberately boring
No Kubernetes, no vector database, no framework-of-the-month. TELOS runs on a standard PHP host with a MySQL database. The sophistication is in the data layer, not the infrastructure bill:
Every AI answer is assembled from a live context pack (standings, full race classifications, race-control incidents, fresh press headlines) fetched per question and cached. The model summarises evidence; it doesn't invent.
Every upstream call (timing API, standings, news, LLM) flows through a caching layer with sensible TTLs. Result: thousands of page views cost cents, and rate limits never become outages.
Model fallback chains, retry with backoff, best-effort context blocks. Any single upstream can fail and the product stays up; it just gets slightly less clever.
Every GP, driver, season, recap and even every chatbot question is a server-rendered, canonical, structured-data-rich URL. Distribution is built into the data model, not bolted on.
The accountability loop
Before every Grand Prix, the TELOS model writes its podium prediction to the database with a timestamp. After the race, an autopilot fills in what actually happened. The result is a growing, irreproducible dataset of AI predictions versus reality, published in full on the accuracy page. No human tipster exposes themselves like this; that asymmetry is the product's moat.
The same philosophy runs through the Race Engineer chatbot (every answer cites clickable sources), the title-race calculator (pure math, no vibes) and the Strategy Sandbox (your pit calls, simulated against the real race).
What shipped
All of it is live and public. Open any tile.
Prediction enginePre-race podium picks with a public, season-long track record.
Open the record Race Engineer AIGrounded F1 chatbot with live web search, cited sources and memory.
Beat TELOSUsers pick podiums against the AI; a leaderboard keeps score.
Strategy SandboxPit-stop simulator on real stint and degradation data.
Embeddable widgetsStandings and prediction widgets any site can embed.
Full SaaS plumbingAuth, Stripe billing, plan gates, push alerts, GDPR consent, observability.
Want this build discipline on your product?
Connexa Studios designs and ships AI products end-to-end: grounding pipelines, accountable models, programmatic SEO, lean infrastructure. TELOS is the live demo; your industry is next.