Research · Build · Deploy

Madusanka
Premaratne

Applied AI Researcher

Building AI systems grounded in research — privacy-preserving LLMs, semantic generalization, and on-device agentic architectures.

Privacy-First LLM Systems

Two research programs — SemanticGuard and EdgeTal — test whether privacy-preserving architecture can hold up against cloud-scale utility. What follows is what's actually implemented, not a generic checklist.

Query-Time Semantic Generalization

SemanticGuard's core method: sensitive fields are generalized at the moment of the query — before anything reaches a cloud LLM — rather than redacted after the fact. Evaluated on 32,593 queries from the OULAD dataset at 95.1% institutional data protection with 93.4% utility preserved.

On-Device Agentic RAG

EdgeTal's full retrieval, reasoning, and generation pipeline runs on local hardware. Zero candidate or user data is transmitted to any cloud inference endpoint — the architecture, not a policy, is what enforces this.

Data Minimization by Architecture

Collecting less, and generalizing what is collected, beats collecting everything and relying on access controls or post-hoc redaction to contain it.

Utility-Preserving, Not Just Private

Privacy techniques that gut model usefulness don't get adopted. The research question is the tradeoff itself — SemanticGuard measures both protection and utility, not protection alone.

Edge-First Inference

Quantized models running on commodity hardware eliminate the round-trip to a third-party API entirely — no data residency ambiguity, no inference-time exposure to a cloud provider.

GDPR-Class Compliance by Design

When raw data never leaves the device, there's no cross-border transfer and no third-party sub-processor to disclose. Compliance follows from the architecture, not a checkbox added afterward.

Why This Matters

Data Leakage via Training

LLMs can memorize and resurface fragments of the data they were trained or fine-tuned on. Without an architectural safeguard, sensitive inputs carry a real statistical extraction risk — not a hypothetical one.

Regulatory Exposure

GDPR/CCPA-class regimes attach real liability to cross-border transfer and undisclosed retention. Architecture that avoids transmission avoids the exposure by design, not by a policy document.

Trust & Adoption

Enterprises evaluating AI vendors increasingly ask where their data goes before what the model can do. An architecture that answers that cleanly is a competitive advantage, not just a compliance line item.

Want the research or a build that applies it to your product?

What I Build

Services grounded in active research — not packaged templates. Every engagement starts with understanding the real problem.

Research-backed

AI Privacy & Compliance Consulting

Design and implement query-time privacy frameworks for LLM applications — semantic generalization, data minimization, and GDPR-compliant AI pipelines without sacrificing utility.

PythonLangChainDifferential Privacy Frameworks+4 more
Learn More
Production-grade

AI Knowledge & Retrieval Systems

Build multi-agent retrieval-augmented generation systems for enterprise use — from architecture design through production deployment, including on-device and edge-first configurations.

LlamaIndexLangChainPostgreSQL (pgvector)+5 more
Learn More
Custom-built

AI Search & Text Intelligence Consulting

Apply semantic generalization, embedding strategies, and NLP pipelines to solve real classification, extraction, and search problems — grounded in current research, not off-the-shelf demos.

Hugging Face (Transformers)spaCyScikit-learn+5 more
Learn More
Privacy-first

On-Device & Data-Sovereign AI Consulting

Move inference off the cloud — design on-device agentic systems that preserve data sovereignty, reduce latency, and achieve inherent compliance. No third-party data exposure.

Ollama (local runs)Llama.cppONNX Runtime+4 more
Learn More
Strategic

AI Strategy Consulting

Define your AI research agenda, evaluate model capabilities, and translate academic findings into product strategy. Research collaborations, manuscript preparation, and AI readiness assessments.

Google Vertex AIOpenAI / Anthropic APIsWeights & Biases (eval tracking)+3 more
Learn More
End-to-end

Enterprise AI Integration

Integrate LLM inference pipelines and multi-agent systems into existing products — from API design and prompt engineering through evaluation, monitoring, and responsible deployment.

FastAPI / Express APIsLangfuse / Phoenix (LLM observability)Docker+4 more
Learn More

Sovereign & Enterprise Tech Stack

We leverage production-grade platforms and open-source models to build secure, private, and high-performance AI solutions.

Python
LangChain
LlamaIndex
HuggingFace
Ollama
PyTorch
Docker
Postgres
10+
Years in Industry
4
Active Research Projects
2
University Collaborations
95.1%
Privacy Preservation Rate
Applied AI Research & Founder

Bridging the gap between Academic Research and Enterprise Application

I am a Founder, CEO, and applied AI researcher with over a decade in enterprise software. I lead the research agenda at Knivok, focusing on privacy-preserving LLMs, query-time semantic generalization, and on-device agentic architectures.

My work translates theoretical models into production-ready software systems for clients globally, backed by first-class academic credentials from the University of Jaffna and London Metropolitan University.

Active Research Collaborations

Applied AI Publications & Experiments

2025 – 2026Active collaboration

EdgeTal

Lead Researcher & Corresponding Author

On-device agentic RAG system for privacy-preserving talent discovery — full retrieval and inference pipeline at the edge, zero candidate data transmitted to the cloud. Presented at EICON 2026 (ESOFT International Conference), Kandy — awarded Best Poster Presenter.

Zero
Candidate data transmitted to cloud services
On-Device
Full agentic RAG pipeline — retrieval, reasoning, generation — at the edge
2025 – 2026Active collaboration

From Surveys to Signals

Corresponding Author & Lead Researcher

An attribution-stability audit of machine-learning driver rankings on correlated HR survey data — testing whether SHAP-based explanations can be trusted, and building a deployable two-item scoring instrument from what actually holds up under resampling.

761
Combined respondents across both survey datasets
96–100%
Full-model accuracy retained using a 2-construct minimal instrument