BLOCKORTALK

Outsmart the scam before the scam outsmarts you.

Inspiration

A caller knows your name. They mention a recent purchase. Then they tell you your account is at risk—and you need to act now.

Would you know what to do?

Scam education often arrives as a warning to read or a quiz to pass. Real scams arrive as conversations. They build trust, introduce pressure, and turn an ordinary interaction into a consequential decision.

We built BLOCKORTALK to give people somewhere to practice that decision before real money or personal information is at stake.

What it does

BLOCKORTALK is an interactive scam-prevention game where players speak with AI callers, assess suspicious requests, and learn from their own responses.

Not every caller is a scammer. Players encounter both legitimate and fraudulent scenarios, so winning requires judgment—not automatically distrusting everyone.

During a call, players can ask questions, challenge a claim, protect fictional secrets, and choose how to respond. Afterward, they give a verdict: Scam likely, Not scam, or Unsure.

Correct verdicts earn 25 BLOKBUKS, incorrect verdicts lose 25, and choosing Unsure carries no penalty. Scored calls update the leaderboard, while completed calls contribute to a personalized profile showing strengths, commonly missed tactics, and suggested practice.

The experience also includes:

Live voice conversations that let players rehearse responding under pressure.
Scenario context and fictional details that make each interaction understandable without requiring real credentials.
Personalized coaching grounded in completed calls and observed decisions.
Skill logs and progress tracking that help players identify what to practice next.
Consumer alerts that connect scam awareness with relevant practice.
Anonymous, aggregated carrier insights that show where training participants struggle across categories and tactics.

For a carrier such as T-Mobile, those aggregate patterns could help answer a practical question: Which safeguards and educational messages do customers need most?

How we built it

We built BLOCKORTALK around three specialized AI responsibilities, supported by NVIDIA Nemotron, shared call state, and persistent records.

Planner Agent

The Planner prepares practice scenarios and recommends exercises using scenario data and available training context. It gives the experience structure: a claimed identity, a plausible situation, a requested action, and a learning objective.

Detector Agent

The Detector examines conversation evidence and scam tactics. It supports the distinction between recognizing a suspicious caller and understanding what made the request risky.

Debrief Agent

The Debrief turns call evidence into explanations and coaching. It helps players understand what happened, what they handled well, and which safer action they could try next.

Across the system, we maintain an important boundary: models interpret and explain; application code controls point awards and stored totals. Rewards should follow explicit rules, rather than an unpredictable model-generated score.

Our stack combines:

NVIDIA Nemotron for planning, analysis, and coaching workflows.
ElevenLabs for the live conversational voice experience.
Photon for the phone-calling integration.
TigerData PostgreSQL for scenario data, call records, player progress, scores, and analysis.
Next.js with Cloudflare’s vinext adapter, deployed on Cloudflare Workers.

This gives Nemotron work beyond answering a chat prompt: preparing exercises, interpreting evidence, supporting recommendations, and producing explanations within a larger application workflow.

Challenges we ran into

Making conversations believable without making the exercise misleading.
A convincing scenario needs context and natural responses, but it must remain a clearly identified training experience using fictional information.

Teaching judgment instead of blanket suspicion.
If every caller is fraudulent, players quickly learn the shortcut. Including legitimate calls makes independent verification—and avoiding false accusations—part of the lesson.

Connecting the entire call lifecycle.
Starting a voice session is only one step. We also needed to carry a completed verdict into persistent history, update scores reliably, refresh profile analysis, and prevent repeated submissions from awarding points twice.

Keeping feedback consistent with the evidence.
A confident explanation is not enough. Coaching needs to reflect what the player actually said and did, while scoring remains predictable and inspectable.

Separating personal coaching from company insights.
Players need individual feedback. A carrier dashboard needs aggregate patterns. Those are different views with different data requirements.

Accomplishments that we’re proud of

We brought voice simulation, competitive progress, personalized learning, and aggregate training insights into one connected experience.

A player can enter a fictional situation, speak with an AI caller, make a decision, see the evidence, and carry that result into future practice.

We are especially proud of making uncertainty a valid response. Choosing Unsure does not cost points: pausing to verify should be a habit the product encourages.

We also built a foundation for evaluating both sides of scam judgment—catching fraudulent callers and recognizing legitimate ones.

What we learned

Recognizing a scam is not the same as responding safely.

Someone can correctly suspect fraud and still disclose a verification code. Someone else can reject a legitimate caller without checking through a trusted channel. Useful coaching must examine the behavior behind the verdict.

We also learned that specialized agents work best with clear responsibilities, shared evidence, and explicit limits. A model can help interpret a conversation; deterministic code can make the resulting reward consistent.

Finally, engagement and effectiveness are separate questions. A leaderboard may motivate practice, but demonstrating safer behavior in unfamiliar situations requires testing beyond the game itself.

What’s next for BLOCKORTALK

We want to make practice more relevant and measure whether the lessons transfer:

Adaptive exercises that target each player’s recurring mistakes.
Seasonal scenarios covering shopping, travel, taxes, and other familiar situations.
Multilingual conversations that make training accessible to more communities.
Stronger conversation-level coaching with clear evidence for each recommendation.
Carrier partnerships to explore training distribution and potential rewards.
Evaluation on unfamiliar scenarios to test whether players choose safer actions after practicing.

Real-world prizes and carrier integrations are future partnership opportunities, not promised benefits. Our next milestone is evidence: whether players can apply what they learn when the caller, story, and pressure are new.

The best time to practice handling a scam is before a real scammer calls.