Project proposal for Axelera AI's The Prompt Project Challenge
In September 2025, a telephone scam in my local community on Evia, Greece, began with an ordinary call to a family's landline. The callers reached a minor while her parents were away. According to the report, they knew personal details and convinced her to gather family jewellery worth thousands of euros for a physical handover.
The case exposes a broader problem. Older people remain especially vulnerable, but anyone can be isolated and rushed into acting before they verify a convincing story. The attack did not require a malicious website or smartphone application. It relied on personal context, urgency, trust, and uninterrupted access to someone through the household telephone.
This pattern is current and organized. In May 2026, the Hellenic Police reported dismantling a group linked to 33 telephone fraud and theft cases and €273,500 in proceeds. Callers impersonated electricity-company staff, warned of an imminent explosion, and instructed victims to leave cash and valuables near the door. They kept each victim occupied on the telephone until an accomplice arrived to collect the property.
The pattern persists despite arrests, with closely similar telephone scams repeatedly reported across Greece.
A Safe Pause for Telephone Scams is a local AI checkpoint for user-authorized landline calls. It detects combinations of social-engineering tactics and high-risk requested actions, then issues a calm warning supported by specific evidence from the conversation. It does not claim to determine that a caller is fraudulent. Its purpose is to interrupt a dangerous request, explain the concern, and prompt independent verification before the user acts.
Possible scam: pause now. “The caller wants cash or valuables collected from your home. Do not hand anything over. End the call and contact someone you trust now.”
Why landlines?
Current scam-protection features mostly target supported smartphones, carrier services, or cloud call forwarding. They do not provide a general, open solution for an existing household landline. Ordinary mobile applications also cannot reliably capture both sides of a native cellular call.
The four-week prototype therefore targets one reproducible landline configuration: a corded telephone with a removable handset cable. A documented Retell 157 passive connector will sit between the telephone base and handset. Its 3.5 mm output will feed the Dell workstation's integrated Realtek audio input through a passive headset adapter if required. Because Dell documents a microphone-capable global headset jack, the baseline needs no PBX, SIP service, telephony card, cloud forwarding, or separate sound card.
The target incremental capture-hardware budget is below €100, excluding the supplied competition computer. Safe Pause will neither answer nor route calls, so a PC or model failure should stop analysis without interrupting telephone service.
The demonstration will support one documented English-language combination of telephone, connector, and audio input. DECT, arbitrary-phone compatibility, line-side integration, Greek, and other languages remain future work rather than dependencies for the initial prototype.
How the system works
fixed service -> selected corded telephone base
|
Retell 157 tap
/ \
telephone handset 3.5 mm audio
|
Dell integrated audio input
|
v
+-------------------------------------------------------+
| Dell Pro Slim Plus XE5 + 16 GB Metis PCIe |
| |
| Host CPU |
| audio -> voice activity -> streaming speech-to-text |
| -> tactic/action evidence -> temporal case |
| | |
| bounded transcript windows |
| | |
| Metis AIPU |
| supported 2–3B SLM -> contextual tactic analysis |
| | |
| Host safety policy |
| validate evidence -> keep checking / warn once |
+-------------------------------------------------------+
|
local display and/or speaker
The host CPU will handle audio capture, voice activity detection, streaming speech-to-text, and deterministic policy logic. The current Voyager Model Zoo does not provide a documented, low-risk Metis pipeline for these dynamic speech workloads. Metis will instead run a supported, precompiled small language model (SLM) for bounded contextual analysis.
The final SLM will be selected from the documented Voyager options using measured latency, memory use, evidence grounding, and output validity. It will receive a bounded multi-turn window of the conversation: a fixed instruction block, the retained call opening where identity and purpose are introduced, and the most recent transcript turns. From that window it will report a small, strictly validated vocabulary of semantic observations covering:
- the identity or authority claimed by the caller;
- urgency, threats, secrecy, or resistance to independent verification;
- requests for cash, jewellery, unusual payment methods, passwords, one-time codes, or remote-access software;
- the action the user is being asked to take; and
- short transcript spans supporting each observation.
The SLM will identify combinations, contradictions, and missing information that isolated keywords cannot capture. A legitimate caller may mention money or electricity. The higher-risk pattern combines claimed authority, a manufactured emergency, resistance to independent verification, and a request for an irreversible action.
The SLM will not choose the warning behavior. A deterministic host policy will validate its output against accumulated conversation evidence and issue at most one decisive warning per session, because repeated warnings train people to ignore them. If the Metis service fails, Safe Pause will say plainly that assessment is unavailable rather than pretend otherwise; the telephone itself is unaffected either way.
This selection-stage proposal intentionally omits exact prompts, thresholds, rule combinations, confidence calibration, and unpublished evaluation cases. The final submission will disclose the materials required for reproduction.
Privacy and safety by design
Telephone calls contain sensitive personal information, so the prototype will operate without a cloud dependency:
- audio, transcripts, model inputs, and warnings remain on the Dell system
- raw audio and full transcript buffers are deleted when each session ends
- test-data capture is a separate, explicit mode limited to consented or synthetic scenarios
- keypad tones will not be intentionally decoded or retained
- the system will never display a green “safe caller” verdict
- and it will not automatically disconnect a call or contact police, a bank, or a family member.
The demonstration will use scripted, synthetic, or informed-consent calls. Safe Pause is an assistive checkpoint, not a caller-authentication system, fraud adjudicator, or emergency service; the user retains control of the call and any response.
What is different about this proposal?
Related protections are emerging but remain limited: Google restricts on-device detection to selected Pixel devices and countries, while LovShield advertises a proprietary subscription for mobile and landline calls. Our contribution is a specific open and reproducible combination:
- Landline-first, local operation: it protects a channel that cannot install an application, without forwarding private call content to a cloud telephony or inference service.
- Evidence-linked, safety-bounded warnings: it identifies the observed tactic and requested action instead of presenting an unexplained risk score. Deterministic policy controls intervention; the SLM cannot declare a caller safe or produce a severe warning without grounded evidence.
- Low-cost, fail-open integration: it reuses the supplied Dell's audio hardware and keeps the telephone outside the AI call-control path.
- Reproducible evaluation: the completed project will release the required Wingman prompt journey, software, configuration, bill of materials, evaluation procedure, and setup instructions. Legitimate bank, utility, medical, family, sales, and support calls will serve as hard-negative tests.
Voyager Wingman and Metis
Voyager Wingman will support the construction, testing, and iteration of the Metis deployment and its integration with the host application. It will be used to:
- configure and validate the supported SLM profile;
- construct the bounded request/response integration;
- diagnose device, runtime, memory, and latency behavior;
- test structured-output validation and failure handling;
- measure SLM inference while the host processes a live audio stream; and
- document prompts, generated artifacts, corrections, and benchmarks.
Metis has a specific application role: contextual analysis of the structured case. The competition Dell's single 16 GB Metis AIPU will run the selected supported 2–3B SLM while the 20-core host processor handles streaming audio. This split keeps each workload tied directly to the use case and to the platform's documented capabilities.
Why I can deliver it
I am a computer science graduate and open-source developer. During a research internship, I evaluated ML models, quantization strategies, and deployment frameworks for edge hardware, balancing accuracy, latency, and energy. I have since applied that in my own projects: quantized vision inference on ARM NPU devices with an Embedded Linux cross-compilation pipeline, and a bachelor's thesis that combined OCR, retrieval-augmented generation, and text-to-speech into one study-assistant system. I have also led two award-winning hackathon teams, delivering working prototypes under hard deadlines.
Four-week delivery roadmap
| Week | Demonstrable outcome |
|---|---|
| 1 — Validate the input path | Establish reliable two-sided telephone audio and timestamped streaming transcription on the supplied platform. |
| 2 — Complete the safety baseline | Implement bounded transcript context, the deterministic evidence-and-warning policy, and controlled scam and benign scenarios, with output validation and failure handling frozen before model integration. |
| 3 — Integrate Metis reasoning | Run the selected supported SLM on Metis under its validated output contract, and measure output validity, latency, and concurrent host/Metis operation. |
| 4 — Evaluate and publish | Run the held-out evaluation, record the offline demonstration, and publish the code, prompts, configuration, parts list, results, limitations, and reproduction instructions. |
This sequence establishes a testable baseline before Metis integration, isolates the highest-risk dependencies early, and leaves the final week for measured evaluation rather than unfinished integration.
Demonstration target
The live demonstration will use the documented landline setup for five controlled scenarios:
- a false electricity emergency requesting valuables;
- a family emergency requesting an unusual payment;
- an impersonated bank employee requesting a one-time code;
- a legitimate utility or bank conversation containing superficially similar words; and
- a normal urgent family call that must not cause a severe scam warning.
Each warning will show the supporting phrase or event and an independent next step. The demonstration will then disable the Metis reasoning process live, showing that Safe Pause reports itself honestly as unavailable while the telephone call continues completely unaffected.
Prototype scope
To remain achievable within four weeks, the prototype focuses on landline conversation analysis and does not depend on smartphone features, caller authentication, automatic call termination, or banking integrations.
The project has one objective:
Recognize when a caller is combining pressure with a dangerous requested action, and help the person pause before the loss occurs.
I would value community feedback on two questions: Which warning wording creates a pause without causing panic? Which legitimate calls would provide the most demanding false-positive tests?
References
The Evia landline incident involving a minor, eVima, September 2025.
A July 2026 telephone-fraud incident in Evia, eVima, July 2026.
Hellenic Police: 33 telephone-fraud and theft cases, €273,500 in proceeds, 9 May 2026.
