Agentic AI Systems: Developed LLM agentic systems, leveraging harness and context engineering for enabling tools, loops and various task executions. Introduced knowledge and skills modules as memory for agents to reference on task execution. Also involved RAG by wiring up relevant integrations and searches (including web) to further widen agent's awareness where relevant.

AI Assistant Communications: Enabled the omni-channel communications infrastructure powering the AI assistant — voice calls, multi-channel messaging (Slack, Microsoft Teams, Discord, WhatsApp), emails, and in-app chat/calls — integrating Twilio, LiveKit, platform bot frameworks and third-party APIs.

Live Meeting Attendance: Built the assistant's ability to join Google Meet and Microsoft Teams calls as a real participant — participant tracking, two-way screenshare, in-meeting chat and auto-leave — via a Recall.ai bridge.

Voice Identity & Call Recording: Built speaker diarization with voice-fingerprint labelling, splitting co-located speakers sharing a single microphone through per-utterance embedding clustering, alongside call recording/egress and speech-aligned transcripts with in-app playback.

AI Desktop & Browser Control: Built a vision-based (VLM) computer-use agent that automates tasks across remote Ubuntu, Windows and macOS desktops, with CLI-installable clients for connecting a user's own machine.

Organisation Support: Built multi-tenant organisation features on top of personal workspaces — member management and invites, role-based access control with custom roles, ownership transfer, and project/resource sharing.

End-to-end model development and deployment, such as data analysis, model training, hyperparameters tuning, model selection and cloud deployment on Azure.

Worked with linear models, trees, random forest, XG boost and light gradient boosting machine (LGBM).

Pipeline External Corrosion: ML. Built pipeline external corrosion depth & length prediction models of >90% accuracy.

Pipeline Freespan Formulation: ML. Built subsea pipeline height & length prediction models of >70% accuracy.

HIP: Computer Vision. Worked on working person detection through object detection with YOLO, pose estimation and vLLM with GPT4o. Involved prompt engineering for better inference performance.

Proposed a typed harness state model with integrity labels, so decay and consolidation apply only to knowledge and skills, while permissions, denies and revocations change only through provenance-backed events.

Built a dual benchmark and plugin layer that scores any LLM and memory backend (including Mem0) on utility under decay and authority-state fidelity over long horizons.

Ran 27 live experiments across four model families and two domains: type-blind memory executed revoked skills 39-97% of the time, versus zero false authority under the typed model at comparable utility (±0.05).

Review Paper. Studied existing tech related to image and music generations, majority being text-conditioned.

Overview Paper. Developed POCs for Diffusion based Music-conditioned Image Generative AI and Transformer based Image-conditioned Music Generative AI.

Applied PageRank and HITS for analysing key driver in the evolution network.

Analysed the community structure of the network with modularity.