Questions I'm Chasing
Diving into the unknowns of AI, multimodal systems, and human-machine interaction. Some questions have answers. Most lead to better questions.
STARC : Signal Triggered Adaptive Refinement Controller
An inference time adaptive self refinement controller for locally runnable large language models
Developing a novel adaptive self-refinement controller that dynamically adjusts the inference process of locally deployable large language models (LLMs) based on real-time signal triggers. This research aims to enhance the efficiency and accuracy of LLMs in low-resource environments, enabling more effective deployment in edge devices and offline scenarios.
Progress
Tags
- Python
- Self-Refinement
- Engineering
- LLM
- Edge AI
- Adaptive Systems
A dual-phase financial strategy for sustainable customer loyalty
Integrating fixed deposits and voucher systems in Sri Lankan supermarket retail
Develops and evaluates an innovative dual-phase customer loyalty strategy for Sri Lankan supermarkets, integrating fixed deposit investments with voucher-based reward systems to enhance both customer engagement and sustainable business profitability during festive seasons. The research employs a mixed-methods approach combining financial modelling using the EBITDA framework, customer surveys (n = 250), and a hypothetical case study with 1,000 participants. The dual-phase model generated projected returns of Rs. 19,135,035.44 from fixed deposits (6.3% gross profit margin) and a Rs. 1,113 markup profit per voucher participant, resulting in an overall EBITDA gross profit of Rs. 2,248,035.44.
Progress
- Finance
- Retail Management
- Customer Loyalty
- Fixed Deposits
- EBITDA
- Sri Lanka
Stochastic Interest Rates and a Dual-Vehicle Savings Structure
A simulation-based retirement planning model for the Sri Lankan private sector
Develops a general framework for retirement planning in Sri Lanka's private sector using mean-reverting Vasicek processes fitted to 26 years of local rate and inflation data, combined with a dual-vehicle structure that locks the accumulated corpus into fixed deposits at each five-year boundary while new contributions accrue in a liquid savings track. A 10,000-path Monte Carlo simulation shows that inflation, rather than the fixed-deposit rate, is the binding constraint on real wealth: the voluntary corpus loses real value in roughly a quarter of simulated futures despite being nominally certain to grow. Abstract submitted to the SLAAS 82nd Symposium (Interdisciplinary and Multidisciplinary Studies track), with a companion mobile app under development.
Progress
- Finance
- Monte Carlo
- Vasicek Model
- Simulation
- Retirement Planning
- Sri Lanka