Vihanga – LLM, AI agent development, Machine learning

A next-generation intelligent assistant platform: a companion app whose intelligence runs fully on-device on iPhone, so user data never leaves the phone. The product combines LLM reasoning, knowledge-graph-based retrieval, semantic clustering, and memory systems under strict privacy requirements and a regulated healthcare pathway. I was the AI architect for the platform and the sole developer of its intelligence layer, working directly with the founders and a clinical domain expert from first discovery through to production.

Client and stakeholders: worked directly with a non-technical founding team, one from finance and one running operations, plus clinical domain experts and one application developer, as the AI architect and sole developer of the intelligence layer, from first discovery through to production.

Business problem and ambiguity: turned an ambiguous product vision, expert guidance delivered as an intelligent companion under a regulated healthcare pathway, into concrete requirements: established what regulatory scrutiny actually meant, who signs off, and that the model could never be the source of content, which decided the architecture. Distilled what a good response looks like, since a companion has no fixed right answer, into an explicit rubric with the clinical experts and measured against it from the first iteration.

End user: built for people having personal, privacy-sensitive conversations with a companion app on iPhone, mental health as the first domain, with the intelligence running fully on device so user data never leaves the phone.

Solution built: designed and built the full intelligence layer: expert-approved source material compiled into a structured knowledge base so responses stayed grounded in verified knowledge rather than the model improvising, retrieval-augmented generation (RAG) and a custom memory layer running in parallel, agent orchestration on the on-device model (CoreML, Apple Foundation Models, MLX), and an evaluation loop with an LLM judge scoring against the clinical rubric, a written reason attached to every score. Also built and ran the team's internal evaluation and metrics platform, frontend and backend, for analysing how the system behaved in production.

Measurable outcome: shipped to production and held a 10 second response budget on device, graded 20,000+ conversation turns through the evaluation loop on live traffic, and ran a regression gate on every change while the platform model updated monthly. Localised and fixed a live retrieval slowdown that real users surfaced and test queries never did. Delivered semantic clustering across 50,000+ records with sub-second inference, and absorbed a brand new vendor foundation model announced mid project as a component swap.

A real-time 3D virtual world platform where AI-driven characters interact live with users: contextual dialogue grounded in domain knowledge, role-based character behavior, and procedural world building. The platform's AI systems were demonstrated at CES 2024 and at an NFL Super Bowl-affiliated Gamerfest event as part of successful investor and public showcases. I built the reinforcement learning agents and the LLM content generation systems behind it, in a role that ran across AI and full stack development as priorities shifted.

Client and stakeholders: embedded for over two years with the technical founder and CEO of a US AI startup and their small engineering team, as the AI and real time graphics specialist, wearing many hats across AI and full stack development as priorities shifted.

Business problem and ambiguity: took on the client's core business problem, proving the platform through live public and investor showcases, which demanded AI-driven characters convincing enough to run in real time 3D; priorities moved constantly with the fundraising and showcase calendar, repeatedly cutting scope down to what could be shown working reliably on a fixed event date.

End user: built for attendees interacting live with the AI characters at major public events, for platform users, and for investors evaluating the platform through those demos.

Solution built: designed and implemented reinforcement learning agents and LLM content generation systems for the GURU virtual world platform, handling real time interaction and procedural world building, with RAG pipelines feeding domain-specific knowledge bases into the characters so dialogue stayed contextual and grounded across scenarios.

Measurable outcome: demonstrated the systems live at CES 2024 and at an NFL Super Bowl affiliated Gamerfest event, contributing directly to successful investor and public showcases at two major industry exhibitions, on an engagement that ran over two years.

A commercial mobile app on the App Store with personalized wine recommendations at the core of the product. The recommendation engine serves personalized results to end users at scale, built from sparse user signals and multi-source catalog data with inconsistent schemas. I built the engine end-to-end, from the first proof of concept through to the production API serving the app.

Client and stakeholders: consulted directly with a non-technical founder building a commercial consumer app, translating business goals into the technical roadmap, and owned the engine end to end as sole developer from the first proof of concept to the production API.

Business problem and ambiguity: solved personalised recommendation from sparse user signals and multi-source catalogue data with inconsistent schemas, where the messy data was the core problem: reconciled product data from sources that disagreed on structure and quality before any model could work.

End user: served consumers choosing wine inside the App Store app, with personalised recommendations as a core product feature.

Solution built: built an LLM-augmented deep learning recommendation engine, from multi-source data aggregation and feature engineering through model training and evaluation, deployed as a production API, with heavy modelling re-run daily in a batch container on AWS while the live path served precomputed per-user results in seconds.

Measurable outcome: kept the system in production for over three years inside a commercial mobile app serving end users at scale, tuned against live results, with distribution monitoring that caught a data-quality incident within a week for a clean rollback.

A reinforcement learning framework for training agents that animate lifelike virtual human characters in real time, developed as my PhD research. It is designed for game developers and researchers working in agent-based control and data-driven animation, and the project page carries renders, papers, and the technical story.

Client and stakeholders: developed as my PhD research rather than a client engagement, working with academic supervisors and peer reviewers, and released as an open source framework so the work serves a real user community beyond the publications.

Business problem and ambiguity: tackled an open research problem with no established recipe, lifelike real time character animation normally demanding heavy motion capture data and long training runs, where defining what to build was most of the work.

End user: built for researchers in agent-based control and data-driven animation, and for game developers and animators, including indie teams without the budget for heavy motion capture pipelines.

Solution built: built a model-based reinforcement learning framework using learned latent dynamics models, with agents operating frame by frame to portray social behaviours and conversational gestures, enabling live animation synthesis from arbitrary voice input.

Measurable outcome: significantly reduced training time and motion capture data requirements compared to state of the art methods, published peer-reviewed work at venues including ICANN, and released the framework as open source so the results are usable directly rather than only through the papers.