Experience

Industry-sponsored research project funded by Xebia IT Architects: “A Privacy-Preserving Framework for Secure Knowledge Transfer in Large Language Models for Cyber Threat Intelligence”.

• I own the vision-plus-text security track. It tests how vision-language models respond to direct and indirect prompt injection and to jailbreak attempts delivered through images and text.
• I built the attack taxonomy from the literature and curated attack data from three public benchmarks (MM-SafetyBench, JailBreakV-28K and VLGuard). Each model is scored by attack success rate.
• The wider project studies privacy-preserving language models, including federated learning, reducing the accuracy lost to privacy noise, and testing against real privacy attacks.

• Prompt injection
• Jailbreaks
• Vision-language models
• Attack success rate

• Flick Network is an accredited e-invoicing provider in the UAE and Oman. I build and support integrations that connect clients' ERP systems (SAP S/4HANA, Tally, Zoho, QuickBooks and others) to its platform through REST APIs, connectors and middleware, for clients in the UAE, Saudi Arabia and Oman.
• The platform is built to handle up to 25,000 transactions per second, with internal document processing under 0.3 seconds. Before each go-live I test the client's invoice flow end to end with Node.js, PostgreSQL and Docker-based deployments.
• During launches and the hypercare period afterwards, I trace production failures in APIs, connectors and message flows back to their cause and fix them so clients stay live.

• Node.js
• PostgreSQL
• Docker
• REST APIs
• ERP integration
• Middleware

• I delivered 500+ product demonstrations and technical discussions with enterprise prospects in India, the UAE, Saudi Arabia and Malaysia, explaining API, ERP, middleware and e-invoicing workflows.
• I turned client requirements into solution outlines for the product and implementation teams, and was promoted into the backend engineering team in July 2026.

• Product demos
• Technical discovery
• Solution design

• Multilingual summarisers write fluent text that can still contradict the article. I built a summarisation pipeline on mT5-Large (1.2 billion parameters), fine-tuned on XL-Sum in five languages: English, French, Spanish, Turkish and Swahili.
• An alignment loss keeps an article and its translation close in the encoder, and a reranker picks the candidate summary with the best factual entailment (mDeBERTa-v3) and semantic centrality (MPNet).
• Cross-lingual semantic consistency improved for every language pair, by more than 50% relative on average and by 0.16 for English and Swahili. Expert ratings rose from 3.42 to 4.13 out of 5.
• The work was published as first author in Knowledge-Based Systems, an Elsevier Q1 journal (2026).

• mT5
• XL-Sum
• mDeBERTa-v3
• MPNet
• PyTorch

• I led outreach and event campaigns for technical workshops and hackathons, which raised participation by more than 450 students.
• I secured sponsorship from brands including Wildcraft and Red Bull and ran several event workstreams at once.
• I joined in July 2023 as Marketing Executive and AI/ML Mentor, teaching Python and machine learning to junior students through hands-on projects, and took over as head of the team in June 2024.

• Leadership
• Sponsorships
• Event management
• Mentoring

Two consecutive industry research projects in Samsung's programme for university students, with Agile product teams.

• Read to conversational speech (Aug 2023 – Jun 2024). I led a team of three students building PACC, which adjusts the intonation, stress and rhythm of read speech and resynthesises it with a HiFi-GAN vocoder. Mean opinion score rose from 3.6 to 4.2 and speech quality reached a PESQ of 4.64. We recorded 100 read and 100 conversational samples for training.
• The work was published as first author in Springer LNNS (ICICC 2025), and Samsung awarded a Certificate of Excellence.
• Automated video captions (Aug 2024 – May 2025). I led a team of four students building an app that turns English and Hindi speech into styled captions using Whisper, with aligned subtitle timestamps and sentiment-based effects and emojis. It cut manual review time for video transcription by 50%. I wrote the JSON styling templates and the Flutter interface.

• HiFi-GAN
• Whisper
• Speech synthesis
• Flutter
• Flask

• I built an IP address scanning and monitoring tool for the Range Operations network. It detects duplicate IP and MAC addresses and changes in device status, supporting real-time vulnerability assessment and security monitoring in a mission-critical environment.
• Reworking the scanning workflow and the MySQL queries behind it improved detection accuracy and responsiveness by 25%.
• I added a device status view to the internal dashboard using Java (JDBC), JSP, AJAX, HTML, CSS and Bootstrap, and submitted the technical report “Study on Advanced IP Address Scanning”.

• Java
• MySQL
• JSP
• AJAX
• Network security