Post Link
https://x.com/interceptoraii/status/2069842468963811828?s=20
Please call the number +1 (629) 253-2827 to experience our demo for yourself. Everything was built with love and legitimately. THERE IS NO MOCK SPEECH INVOLVED, DATA IS BEING PROCESSED LIVE.
Inspiration
Interceptor came from something very real to us.
We have seen the damage scam callers can cause, especially to vulnerable elderly people. It is easy to say, “Just do not pick up unknown calls” or “Just be careful.” But in real life, it is not that simple.
Volunteering as a Digital Ambassador with Seniors Go Digital, I've seen first hand the heavy burden seniors carry to separate scams from legitimate calls.
A scam call does not feel like a scam when it is happening. The caller sounds serious. They may claim to be from the police, a bank, or a government agency. They create fear. They make the person feel like they must act quickly. For an elderly person who is alone at home, that pressure can be overwhelming.
That was what inspired us to build Interceptor.
We did not want to build another tool that only warns people after something bad happens. We wanted to build an AI agent that steps in during the call itself, before the scammer reaches the vulnerable person.
What it does
Interceptor is an AI agent that protects people from scam calls.
When an unknown caller contacts a protected user, Interceptor answers first. The user does not need to open an app, read a warning, or decide whether the call is safe. Interceptor handles the caller on their behalf.
The agent asks safe questions to understand who the caller claims to be, why they are calling, and what they want. At the same time, it checks for scam behaviour such as threats, urgency, secrecy, impersonation, and requests for money.
If the call looks suspicious, Interceptor blocks the call and sends a clear alert to the caregiver through Telegram. The alert includes the risk level, evidence, summary of the call, and recommended next action.
The goal is simple. The scammer should not get the chance to pressure the victim directly.
How we built it
We built Interceptor around a multi-agent workflow.
The first agent is the conversational agent. It speaks to the caller, keeps the conversation controlled, and extracts the important details. It does not reveal personal information. It does not confirm household details. It only asks what is needed to understand the caller’s intent.
The second agent is the verification agent. It runs in parallel and checks the call for scam signals. It looks at the caller’s claim, the requested action, the language used, and whether the pattern matches known scam behaviour.
Once both agents have enough information, Interceptor makes a decision. The call can be passed through, reported, or blocked. Caregivers receives a Telegram alert with the key evidence and next steps.
We also built a sample knowledge base flow. When a scam is confirmed, Interceptor can store the scam signature. This helps the agent recognise similar scripts in the future, even if the scammer changes phone numbers.
Challenges we ran into
One challenge was making Interceptor careful but not overly aggressive.
Not every unknown caller is a scammer. Some calls may be from delivery drivers, clinics, banks, or legitimate services. We could not simply block every unknown number. The agent had to look at behaviour, not just caller identity.
Another challenge was deciding how much the agent should ask. If it asks too little, it may not get enough evidence. If it asks too much, the scammer gets more time to manipulate the conversation. We had to keep the questions short, safe, and purposeful.
We also had difficulties with Interceptor's response speed. As the reasoning agent takes a while to process, we decided to use Gemini 2.5 Flash for the response agent while the reasoning agent (ChatGPT 5.4 mini) simultaneously takes in the transcript live to do deeper reasoning. This ensures that the call feels natural with minimized latency.
We also had to think about privacy. The agent should never reveal that the user is elderly, alone, or unavailable. It should not confirm addresses, NRIC numbers, bank details, or family information. A scammer may try to extract these details, so the agent has to protect them by default.
The Telegram alert was another important challenge. We wanted the caregiver to understand the situation quickly. A long technical report would not help. The alert had to answer three things clearly: what happened, how risky it is, and what to do next. At the same time, it needed sufficient evidence for caregivers to trust Interceptor.
Accomplishments that we're proud of
We are proud that Interceptor feels practical.
It is not just a chatbot. It is an AI agent designed to stand between a scammer and a vulnerable person during the most dangerous moment of the scam.
We are also proud of the multi-agent workflow. The conversational agent and verification agent work together in real time. One handles the caller while the other checks whether the call is suspicious. This makes the system stronger than a simple rule-based blocker.
The Telegram output is another part we are proud of. It gives the caregiver a clear summary, risk status, and next action. It turns a confusing phone call into something understandable and actionable.
Most of all, we are proud that Interceptor addresses a real problem. Scam calls are not just an inconvenience. They can destroy savings, trust, and peace of mind. Building something that can protect people before the damage happens is what makes this project meaningful to us.
What we learned
We learned that protecting vulnerable users is not just about building more features.
Sometimes, the best interface is no interface. An elderly person should not have to analyse a scam call under pressure. They should not have to decide whether a caller is real while being threatened. The protection should happen quietly before the scammer gets through.
We also learned that scam detection is more than checking phone numbers. Scammers can change numbers quickly, but they often reuse the same methods. They use authority, fear, urgency, secrecy, and financial pressure. Interceptor focuses on detecting those patterns.
Another big lesson was that AI agents need strong boundaries. The agent must know what not to say. It must not reveal private details, follow unsafe instructions, or get manipulated by the caller.
This project helped us see AI agents not just as tools for convenience, but as tools for protection.
What's next for Interceptor
Next, we want to make Interceptor more accurate and reliable, while being more dynamic.
We see potential in identifying tonality of speech, which we believe can act as additional evidence to increase Interceptor's confidence.
We want to improve the verification agent so it can recognise more scam types and compare calls against a stronger knowledge base. We also want to refine the risk scoring system so Interceptor can better separate real threats from harmless unknown calls.
We also want to give caregivers more control. Some families may want all unknown calls screened. Others may only want high-risk calls blocked. Interceptor should be flexible enough to fit different households.
In the future, Interceptor can also go beyond elderly protection. Scam calls can affect anyone. As scams become more convincing, even younger people can be pressured into making bad decisions.
Our long-term goal is for Interceptor to become an AI agent that protects people before scammers can reach them. Not after the money is gone. Not after the victim is scared. Before the damage starts.
Built With
- elevenlabs
- next.js
- openaiagentsdk
- react
- supabase
- telegramapi
- twilio
- typescript
- vercel