Cybersecurity & Digital Solutions Intern
The four-week Cyshield Egypt internship split each day between cybersecurity and AI & digital solutions. On the cybersecurity side I moved from governance and risk into secure infrastructure, digital forensics, and hands-on penetration testing. On the AI side I covered full-stack development, cloud-native tooling, and applied machine learning, building an end-to-end Whisper-to-GPT-2-to-gTTS speech pipeline as my main project.
Internship cohort celebration
Production Design
The speech pipeline design also considered two production requirements beyond the delivered demo: multilingual interaction and real-time streaming.
For multilingual support, the approach would be a multilingual Whisper variant, language detection on the ASR output, and routing text to an LLM trained for that language, using translation as an intermediate step (ASR, then translate to English, then generate, then translate back) for languages without a strong native-language LLM available.
For low-latency, real-time use, the sequential ASR-then-LLM-then-TTS structure would need to become a streaming one: chunked ASR emitting partial transcriptions as audio arrives, the LLM pre-generating against those partial transcripts, and incremental TTS starting to speak before the LLM finishes its full response, all coordinated through async queues instead of blocking stages.
What I worked on
Cybersecurity track
- Governance, Risk, and Compliance (GRC) foundations: security policies, frameworks, controls, risk management, and compliance reporting
- Digital Forensics and Incident Response (DFIR): evidence acquisition with KAPE and forensic analysis aligned to GIAC GCFA concepts, alongside network and web security fundamentals
- Web application penetration testing through PortSwigger Academy and Burp Suite, covering fuzzing, stored XSS exploitation, and webhook-based attack surfaces
- Network reconnaissance and traffic analysis with Wireshark, plus broader network scanning, vulnerability detection, and exploitation-technique practice
- OverTheWire war games, regex-driven log analysis, and OSP/CH cyber-range exercises for applied investigative problem-solving
- Security Operations Center (SOC) workflow simulations, vulnerability scanning, incident-response procedures, and secure coding practices for identifying and patching flaws
AI & digital-solutions track
- Full-stack development with the MERN stack, alongside mobile prototyping in DartPad (Flutter), Kotlin, and Swift, and exposure to ERP platforms
- Cloud-native fundamentals through Kubernetes (CKA/CKAD-aligned material) and DevOps workflows, including container orchestration and deployment pipelines
- NLP, speech-processing, and computer-vision fundamentals, applied directly to NLP/CV model-building and data-engineering tasks
- Big data and business-intelligence concepts alongside embedded-systems topics: ORAN software development lifecycle, LTE air-interface protocols, and IoT/smart-systems integration
- Rapid prototyping on low-code/no-code platforms (Odoo, Webflow, Squarespace) for portfolio sites and quick builds
Completed Work
Speech-to-speech pipeline
A speech-to-speech pipeline that chains three open-source models in Google Colab: OpenAI's Whisper (whisper-tiny.en) for speech recognition, GPT-2 to generate a text reply, and Google's gTTS to turn the reply back into speech. I also handled model loading, GPU/CPU device configuration, sample data preparation with Hugging Face's datasets library, and debugging dependency and memory issues.
- Built with
- Whisper (whisper-tiny.en), GPT-2, gTTS, Hugging Face Transformers, Google Colab
- Contribution
- Built and integrated all three pipeline stages, including the surrounding device-configuration and debugging work.
- Outcome
- A working speech-in, speech-out conversational demo running entirely on free-tier Colab compute.
Technical Stack
Cybersecurity skills
Digital forensics (KAPE, GIAC GCFA concepts), Web application penetration testing, Network traffic analysis, GRC and SOC workflows, Secure coding practices
AI and software-engineering skills
MERN stack development, Mobile prototyping (Flutter, Kotlin, Swift), Kubernetes and container orchestration, NLP, speech processing, and computer vision, Speech pipeline engineering
Tools and platforms
KAPE, Wireshark, PortSwigger Academy, Burp Suite, OverTheWire, Whisper, GPT-2, gTTS, Hugging Face Transformers, Kubernetes, Docker, Odoo, Webflow, Squarespace
Problems and what I would change
I chose whisper-tiny.en for its speed and low resource usage, but it is less accurate, which matters where transcription errors are costly.
The right model size depends on the accuracy you need and the compute you have, so the largest model is not automatically the answer.
whisper-tiny.en was fine for a proof of concept on free Colab. For production I would use whisper-small or larger with quantisation, batching and mixed precision, or reprocess uncertain cases with a larger model.
The pipeline broke on importing Conversation from transformers: the class had been moved or deprecated in a newer version than the one the code was written against.
Colab's package versions can change between sessions, so a notebook that works today may not import cleanly tomorrow.
I would pin exact versions in a requirements.txt and test in a reproducible Docker environment, so breaking changes in a library like transformers show up before a demo, not during one.
Inference hit CUDA out-of-memory errors, and torch.cuda.empty_cache() only gave short-term relief.
Clearing the cache does not fix the cause. Lasting headroom has to come from the model and infrastructure choices.
A proper fix would be a smaller model variant, int8/fp16 quantisation, splitting the workload across GPUs, and an inference server such as Triton that reuses memory natively. Gradient checkpointing would help if training were involved.
Results
- Programme length
- 4 weeks (Jul 2025)
- Pipeline models chained
- 3 (Whisper, GPT-2, gTTS)