33-TruthGuard

Appearance

Step 2.5: Say It Out Loud (Master Spoken Scripts & Rehearsal Guide) ​

This single rehearsal reference brings together every first-person spoken script across TruthGuard. Use this page to practice speaking naturally during technical interviews, screen shares, and system design rounds.

1. Quick Project Pitches ​

30-Second Elevator Pitch ​

(Target Time: 25–30 seconds | Conversational Pace)

"I built TruthGuard, an integrated cyber-defense and digital media forensics platform designed to detect generative AI deepfakes, surveillance camera tampering, and voice phishing scams. Rather than relying on a single static check, I architected a federated micro-frontend suite connecting five specialized engines—combining client-side EXIF metadata parsing and Web Crypto SHA-256 hashing with serverless edge AI functions for multi-modal verification. The platform unifies image error level analysis, telecommunication scam triage, and cryptographic asset provenance into an NVIDIA-inspired dark workstation. It solves the growing crisis of synthetic media by giving security analysts instant, multi-layered evidence verification with definitive confidence scores."

2-Minute Architecture & Technical Pitch ​

(Target Time: 1:45–2:00 minutes | Technical Walkthrough)

"With the recent explosion of generative AI—such as diffusion image generators, ElevenLabs voice cloning, and automated phishing bots—traditional security perimeters are completely blind to synthetic media. Bad actors can now bypass biometric gates, spoof executive phone calls, and tamper with physical CCTV feeds. I built TruthGuard to give security teams and journalists a single, unified forensics workstation to detect and mitigate these multi-modal threats in real time.

Architecturally, TruthGuard is structured as a federated micro-frontend platform. At its core is a central Hub built with React 18, TypeScript, and Vite. The Hub coordinates five specialized forensic engines:

First, TruthLence, which inspects uploaded imagery, documents, and news URLs for generative compression artifacts, cross-referencing claims against established fact-checking sources through Supabase Edge Functions.

Second, DEEP SCAN, an optical tamper and camera forensics engine. It processes media right in the browser using the Web Crypto API to calculate SHA-256 hashes and exifr to extract detailed camera metadata, GPS coordinates, and detect whether editing software like Photoshop, GIMP, or Canva altered the image.

Third, DeepGuard, which triages telecommunication threats. It analyzes phone call transcripts and SMS messages for voice synthesis patterns, artificial urgency, and smishing attacks.

Fourth, MediaVault, a cryptographic asset custody drive built with Zustand that handles file ingestion, simulated multi-stage encryption, and immutable activity audit logging.

And fifth, the VeritAI Executive Dashboard, a 1,600-line analytical command center featuring radial confidence gauges, source reliability tiers based on Media Bias/Fact Check standards, and emotional sensationalism meters.

One engineering highlight I'm proud of is our client-side dynamic port prober in FeatureEmbed.tsx. It actively probes localhost ports using non-blocking HEAD requests to discover running micro-frontend instances dynamically without requiring a hardcoded backend registry.

Overall, TruthGuard combines client-side cryptographic speed with serverless AI intelligence to provide deep, explainable, and trustworthy forensics against modern synthetic threats."

2. Feature-by-Feature Spoken Scripts ​

F1: TruthLence Deepfake & Multi-Modal Verification ​

(Target Time: 60–75 seconds | Focus: Serverless Edge AI & Fact-Check Grounding)

"In TruthLence, I built the multi-modal content verification engine located in truthlence/src/pages/Index.tsx. Security analysts and journalists frequently encounter suspicious media—whether it's an AI-generated portrait, an altered PDF document, or a viral news link—and need instant verification without complicated setup.

When the user drops a file or pastes a URL, handleFileSelect extracts the text or generates a media descriptor, truncating large inputs to 5,000 characters to prevent API timeouts. I then invoke our Supabase Edge Function analyze-media using the Supabase JavaScript SDK. The edge function executes our heuristic forensic rules, checking for diffusion generation artifacts and cross-referencing claims against PolitiFact and Snopes fact-check feeds. When the payload returns, I generate a unique UUID using crypto.randomUUID(), update our state history, and render AnalysisResult.tsx with an animated confidence bar, verdict badge, and clickable citations.

I chose a serverless Edge Function architecture rather than bundling large machine learning weights directly into the browser. Running deep neural networks clientside would cause immense bundle bloat and crush low-end mobile devices, while an Edge Function keeps the client bundle under 300 kilobytes.

If the network drops or the Edge Function times out, the catch block catches the exception immediately, suppresses crashes, and fires a descriptive toast notification using our useToast hook so the analyst can retry.

One limitation is that large video files currently have their metadata inspected rather than full frame-by-frame neural decoding. In production, I would scale this by queuing video jobs through an asynchronous Redis pipeline with worker nodes."

F2: DEEP SCAN Optical Camera Tamper Guard ​

(Target Time: 60–75 seconds | Focus: Web Crypto SHA-256 & Client-Side EXIF Auditing)

"In DEEP SCAN, located in camara-temper/src/lib/fileAnalyzer.ts, I built our client-side camera tamper and forensic metadata analyzer. In video surveillance and digital evidence, attackers frequently tamper with cameras using physical occlusions like spray paint, or manipulate digital frames in editors like Photoshop before submission.

When an investigator drops a suspect media file, handleFileSelect initiates analysis. First, I use the browser's native Web Crypto API to calculate a cryptographic SHA-256 hash directly from the file buffer, creating an immutable tamper-evident fingerprint. Next, I parse EXIF and TIFF headers using exifr to extract low-level hardware attributes like camera make, lens model, aperture, ISO, and GPS coordinates.

My heuristic algorithm starts at a baseline authenticity score of 75. If camera make and model match genuine hardware, I add 10 points. If original capture dates are stripped, I dock 5 points. Crucially, I inspect the software tags against known raster editors like Photoshop, GIMP, Canva, and Figma. If an editing signature is detected, I dock 20 points and raise a high-severity danger alert.

I chose native client-side Web Crypto and exifr rather than uploading images to a remote server. This allows analysis to complete in milliseconds, preserves evidence chain of custody, and guarantees that sensitive surveillance footage never leaves the investigator's machine.

If a file has corrupted EXIF headers, the parsing routine catches the error silently, falls back to raw binary metrics, and flags that metadata was stripped.

One current limitation is that video optical flow analysis is limited in the browser. If I rebuilt this for an enterprise SOC, I would integrate WebGL canvas shaders to analyze frame-by-frame histogram deltas in real time."

F3: DeepGuard Scam Call & SMS Phishing Classifier ​

(Target Time: 45–60 seconds | Focus: Telecommunications Vishing & Smishing NLP)

"In DeepGuard, located in calldetection-sms-contentdetection/src/pages/, I built our telecommunication fraud and linguistic scam classifier. Phishing has evolved dramatically—attackers now use generative voice cloning to impersonate executives on the phone and automated bots to blast highly tailored SMS smishing links.

Across three specialized interfaces—Call Detection, SMS Detection, and AI Content Detection—the user inputs a transcript or text message. When they click Analyze, the component validates the input and calls our Supabase Edge Function analyze-content, passing the modality type and text payload.

The edge function analyzes the text for psychological pressure patterns, urgency markers, and deceptive wire transfer requests. When the evaluation returns, ResultCard.tsx renders a color-coded verdict card: red for fake, amber for suspicious, and green for legitimate. I implemented Framer Motion to smoothly animate the confidence bar and render a bulleted breakdown of the exact risk indicators found.

I chose a serverless edge function architecture because linguistic threat patterns evolve rapidly. By keeping classification logic in serverless edge functions rather than hardcoding static regexes in the frontend, security teams can update detection models in real time without requiring client redeployments.

If the network fails, the try-catch block catches the exception and displays an actionable error toast while re-enabling the submission button.

One limitation is that transcripts are currently submitted as text. As a future improvement, I would connect a Web Audio API recorder to transcribe incoming audio streams directly in the browser using WebAssembly Whisper before classification."

F4: MediaVault Cryptographic Storage & Custody Chain ​

(Target Time: 45–60 seconds | Focus: Zustand State Management & Custody Audit Trail)

"In MediaVault, located in mediavault 2/src/lib/mediavault-store.ts, I built our cryptographic asset storage and evidence custody drive. Digital forensics requires strict chain-of-custody tracking—every time an investigator handles a file, they need proof of when it was ingested, how it was stored, and whether its contents were modified.

I used Zustand to create a centralized reactive store with local persistence. When an analyst drops files into the dropzone or submits a URL, addFilesToQueue stages them as pending queue items. When they trigger the upload, startUpload executes a multi-stage pipeline: it reads the binary data into Base64 using FileReader, simulates progressive cryptographic staging across 30, 65, and 100 percent completion intervals, generates file metadata, and commits the asset to our library.

Simultaneously, the pipeline logs an immutable entry into our activityLog, tracking the exact file name, timestamp, and action. I wrapped all local storage writes in an exception-safe saveToStorage helper to prevent browser storage quota crashes.

I chose Zustand because it provides an un-opinionated, lightweight state machine without the heavy boilerplate of Redux, while offering instant reactivity across the file library, queue panel, and storage widget.

If a user attempts to upload files larger than the browser's local storage limit, the store gracefully retains the assets in active memory while logging an audit warning.

In a full production deployment, I would connect this front-end ingest pipeline directly to an AWS S3 bucket configured with Object Lock for true WORM—Write Once, Read Many—regulatory compliance."

F5: TruthGuard Unified Hub & VeritAI Threat Dashboard ​

(Target Time: 60–75 seconds | Focus: Dynamic Port Probing & Executive Credibility)

"In the main hub and executive dashboard, located in mainpage/src/pages/Hub.tsx and VeritAIApp.tsx, I built the central orchestration platform for TruthGuard. When dealing with multiple specialized security tools, users get frustrated jumping between disconnected terminal windows or obscure port numbers.

I created a unified portal using React 18 and Tailwind CSS with an NVIDIA-inspired dark glassmorphism design. The hub displays six interactive cards representing our detection modules.

A technical feature I'm particularly proud of is our dynamic port scanner in FeatureEmbed.tsx. When a user chooses a tool, rather than hardcoding static links, the component initiates checkPort starting at port 5176. It issues non-blocking HEAD requests with no-cors mode, stepping through local ports until it receives an active response, automatically detecting which port the tool is running on.

For executive decision-makers, navigating to /dashboard loads VeritAIApp.tsx, an in-depth forensic dashboard. It calculates a composite credibility score from four sub-indexes: factual consistency, source reliability, sensationalism clickbait percentage, and cognitive balance. We cross-reference claims against PolitiFact, Snopes, and Reuters, and integrate Supabase authentication to enforce role-based access for journalists and analysts.

I chose this federated hub pattern over an all-in-one monolith because it allows forensic modules to be developed, scaled, and deployed independently while presenting a seamless, unified interface to the investigator.

If an investigator navigates to a module that isn't running, the scanner times out gracefully and displays clear instructions on how to start the service locally.

With additional time, I would replace local port scanning with an automated service worker gateway and unified single sign-on cookie propagation."

3. Interview Rehearsal Checklist ​

Tick each box once you have explained the feature out loud, without reading notes, within the target time:

  • [ ] 30-Second Elevator Pitch — Delivered under 30 seconds smoothly without pausing.
  • [ ] 2-Minute Platform Walkthrough — Covered problem, architecture, all 5 engines, and port prober in under 2 minutes.
  • [ ] F1: TruthLence Deepfake & Fact-Check — Explained analyze-media edge call, payload slicing, and fact-checking citations.
  • [ ] F2: DEEP SCAN Optical Tamper Guard — Explained Web Crypto SHA-256, exifr camera headers, and photo-editor danger flags.
  • [ ] F3: DeepGuard Scam Call & SMS Phishing — Explained vishing/smishing NLP triage, edge function dispatch, and ResultCard indicators.
  • [ ] F4: MediaVault Cryptographic Drive — Explained Zustand reactive store, multi-stage upload progress, and custody audit logging.
  • [ ] F5: TruthGuard Hub & VeritAI Dashboard — Explained dynamic port scanner (HEAD / no-cors), MBFC credibility scoring, and Supabase Auth.