Sukhraj Singh
All projects

AI research · 2026

SCAMPA / ScamShield

Research into local voice agents for controlled scam defence.

Research proposal + separate demoThree-person research team + individual reconstruction

My contribution

I co-authored the research architecture and built the separate deterministic portfolio reconstruction.

Evidence

IEEE-formatted course report, repository, recorded reconstruction and interactive controlled lab.

Scope & limitations

The reconstruction demonstrates the policy boundary; it is not proof of local ASR or LLM inference.

  • C++20
  • Whisper
  • llama.cpp
  • RtAudio
  • Vulkan
  • CMake

See it in action.

00:41 · Local recording
Controlled voice fixture -> transcript adapter -> deterministic safety policy -> bounded response -> local desktop TTS.
Open the interactive reconstruction
About this demonstration

This is a 1080p screen recording of a separate portfolio reconstruction based on the Phase 1 architecture. It executes real local rule analysis, calculates a risk score, enforces a CLARIFY + STALL action and sends the bounded response to desktop text-to-speech. It is not presented as the deleted Phase 2 showcase or as proof of local ASR or LLM inference.

Controlled voice fixture -> transcript adapter -> deterministic safety policy -> bounded response -> local desktop TTS.

The problem & approach

The problem

The research asks how a real-time local agent could engage scam callers long enough to support controlled cybersecurity study without disclosing sensitive information, executing harmful actions or depending on cloud inference. That requires conversational latency, safety policy and privacy to be designed as one system rather than separate features.

The approach

SCAMPA is represented by the ScamShield-LLM Phase 1 work on my GitHub: an IEEE-formatted course research report and proposed modular architecture for a fully local conversational agent. The design connects speech recognition, policy enforcement, local LLM inference, speech synthesis, audio I/O and performance instrumentation while keeping the research boundary controlled and defensive.

How it works
  1. 01Controlled inbound audio
  2. 02Voice activity detection
  3. 03Whisper-compatible ASR
  4. 04Dialogue state + policy engine
  5. 05Local llama.cpp inference
  6. 06TTS + RtAudio output
  7. 07Optional Vulkan acceleration
  8. 08Metrics + Chrome Trace logs
  • I co-authored the Phase 1 report with Martin Jajarmi and Yahya Fazal for an Information Technology Security and Privacy course project at Algoma University.
  • We defined a controlled, inbound-only research scenario and documented the ethical, privacy and consent boundaries around voice data.
  • We proposed a platform-independent engine core for dialogue state, policy enforcement, prompt construction, inference orchestration and metrics logging.
  • We separated platform-specific speech, audio and GPU dependencies behind narrow adapters so ASR, TTS and LLM backends can be replaced.
  • We selected a local C++20 stack around Whisper-compatible ASR, llama.cpp-compatible inference, RtAudio, GLFW, Dear ImGui and optional Vulkan acceleration.
  • We defined future evaluation around end-to-end latency, ASR accuracy, model response time, safety-policy triggers and sustained engagement time.
  • We documented Phase 2 as planned implementation and controlled validation rather than presenting the proposal as a completed product.
  • For this portfolio, I built a separate deterministic reconstruction that executes the documented safety boundary against a controlled voice fixture and sends its bounded response to local desktop text-to-speech.
Technical decisions & lessons

Local-first processing

The design keeps audio and model inference on the research machine to reduce unnecessary data exposure.

Explicit policy engine

Safety decisions remain outside unconstrained model generation so stall, clarify and refuse actions can be enforced in code.

C++20 core

The proposed static-library boundary supports low-level audio, inference and performance control across desktop platforms.

Narrow adapters

ASR, TTS, audio and GPU implementations can change without rewriting dialogue and policy logic.

Trace-based metrics

The proposal makes latency and policy behavior measurable instead of relying on subjective conversational impressions.

Challenges

  • The architecture has to balance low response latency with safety checks that cannot be skipped when the conversation becomes unpredictable.
  • Voice research introduces consent, privacy, retention and telecommunications constraints, so the design is restricted to controlled testing.
  • Local models, speech processing and optional GPU acceleration create hardware and cross-platform integration constraints that Phase 2 still needs to validate.
  • The current repository contains the report and proposed architecture, not the engine implementation; the case study keeps that boundary explicit.

What I learned

  • I learned to place safety policy in the system architecture rather than treating it as a prompt-writing detail.
  • I learned how ASR, local inference, TTS and audio routing contribute separately to end-to-end conversational latency.
  • I learned why replaceable adapters are important when hardware, models and platform APIs evolve independently.
  • I learned to describe research-stage work by what is defined, what is measurable and what still has to be implemented.

Proposed policy boundary · research model

audio -> ASR -> dialogue state
              -> policy: STALL | CLARIFY | REFUSE
              -> local LLM -> TTS

measure: latency | accuracy | policy triggers