Strategic Impacts. Impacts.
Case studies in AI engineering, data science, and enterprise architecture—focused on solving complex business bottlenecks.
AI Job Hunter: Autonomous Talent Acquisition System
Lead AI Engineer & Architect
The Challenge
The manual process of discovering, matching, and applying for high-signal opportunities is extremely inefficient and prone to low quality. Most automation tools use 'brute force' which results in poor outcomes.
The Solution
Architected a multi-agent system using LangChain and Gemini 3.1. It features a decision layer that scores opportunities before applying, a RAG pipeline for dynamic context injection, and Playwright for human-like browser automation.
The Impact
Built an autonomous system that reduces job application time by 90% through intelligent filtering and RAG-based personalization.
Built an autonomous system that reduces job application time by 90% through intelligent filtering and RAG-based personalization.
Technical Engine
Predictive Analytics for Healthcare Payment Forecasting
Lead Data Scientist / Researcher
The Challenge
Healthcare revenue cycles are often reactive. Organizations struggle with delayed payments and high denial rates because they can't predict which claims are at risk.
The Solution
Developed a proactive framework using Random Forest and Neural Networks on a 3-year dataset (~28k records). Leveraging ML to forecast payment behaviors in RCM, achieving R² > 0.90 for financial prioritization.
The Impact
Enhanced financial predictability for large-scale healthcare providers and achieved R² > 0.90 accuracy.
Leveraging ML to forecast payment behaviors in RCM, achieving R² > 0.90 for financial prioritization.
Technical Engine
RevOps AI: Autonomous Data Auditing Platform
System Architect & Full Stack Developer
The Challenge
Billing errors and data hygiene issues cost healthcare providers millions. Manual auditing is slow, inconsistent, and unscalable.
The Solution
Built a platform with an Autonomous Agent Architecture (CEO, Eng, ML Pods). It uses a specialized logic component to scan for architectural patterns and Scikit-learn models for payment forecasting.
The Impact
Reduced manual audit effort by 40% and improved KPI visibility for financial controllers through real-time predictive dashboards.
A B2B SaaS transforming Healthcare RCM through autonomous data auditing and agentic pods.
Technical Engine
TrainIQ: Multi-Tenant Corporate Learning Platform
Full Stack Architect & Main Developer
The Challenge
Organizations need secure, tenant-isolated corporate training with assessments, proctoring, task management, and SaaS billing — without leaking data between customers or relying on cloud AI APIs.
The Solution
Architected a Flask-based platform with three-layer tenant isolation (tenant-scoped PostgreSQL, per-tenant MongoDB GridFS databases, Redis-backed limits), a local AI suite via Ollama (LearnIQ tutoring, AnalyticsIQ insights, ProctorIQ trust scoring, CreatorIQ course generation), Stripe subscription billing, and a Prometheus/Grafana observability stack.
The Impact
Streamlined training for 50+ team members and evolved into a productized SaaS with trial-to-enterprise plans, 350+ automated tests, and a three-process production topology (web, platform console, ops worker).
A multi-tenant SaaS LMS with courses, timed exams, AI-assisted proctoring, local AI tutoring, Stripe billing, and a CEO operations console.
Technical Engine
CareerForge AI 3.0: Voice-AI Career OS
Full Stack AI Engineer
The Challenge
Job seekers lack realistic, real-time data and high-fidelity practice tools for competitive tech roles.
The Solution
Built a multi-agent career OS featuring low-latency voice interactions (Gemini 3.1 Live API) and real-time market grounding using Google Search integration.
The Impact
Built a voice-first career platform that provides real-time market grounding and interview coaching.
Revolutionizing how candidates prepare for high-stakes interviews with real-time AI feedback.
Technical Engine
ReviewRadar AI: Consumer Sentiment Intelligence
AI & Data Engineer
The Challenge
Businesses struggle to extract actionable insights from massive volumes of unstructured customer feedback.
The Solution
Developed an end-to-end sentiment intelligence platform with multi-threaded ETL pipelines, ensemble ML models, and VADER sentiment scores.
The Impact
Captured and processed 7M+ data points to provide deep technical insight into consumer behavior.
Processing 7M+ reviews to generate actionable sentiment insights for business growth.
Technical Engine
CodeX: Competitive Coding & Algorithmic Practice
Full Stack Developer & Algorithmic Researcher
The Challenge
Traditional coding platforms often lack specialized local-first performance tracking and specialized practice modes.
The Solution
Built a specialized environment focusing on algorithmic efficiency, featuring real-time execution metrics and structural complexity analysis.
The Impact
Developed a high-performance environment for mastering data structures and algorithmic complexity.
A robust platform for high-performance algorithmic practice and competitive coding simulations.
Technical Engine
Disaster Management System: Centralized Response Hub
System Designer & Lead Developer
The Challenge
Coordination during natural disasters is often fragmented, leading to delayed relief and inefficient resource distribution.
The Solution
Developed a centralized response platform integrating real-time reporting, resource tracking, and volunteer coordination modules using PHP and MySQL.
The Impact
Centralized relief efforts and optimized resource distribution for faster disaster response.
Digitizing relief coordination and resource allocation during critical natural disasters.
Technical Engine
Statistical Application: Financial Forecast Modeling
Statistical Researcher
The Challenge
Financial forecasts often fail due to improper validation of underlying statistical assumptions in high-volatility environments.
The Solution
Engineered a specialized application for comparing predictive performance across various statistical models, incorporating sensitivity analysis.
The Impact
Improved forecast reliability through advanced statistical validation and error distribution checks.
Applying rigorous statistical methodologies to validate financial forecasting accuracy.
Technical Engine
Kapruka Flow AI: Agentic Commerce on MCP
AI Engineer & System Architect
The Challenge
Sri Lankan shoppers describe what they want in English, Sinhala, and Tanglish, but every storefront forces them to translate that intent into keywords and then act as their own planner, price optimiser, and delivery validator.
The Solution
Built a deterministic planner over the public Kapruka MCP: a multilingual intent parser produces a structured frame, relevance scoring and budget-constrained search assemble Ideal, Cheaper, Premium, and Fast carts, and every tool call is surfaced in a live activity feed. Next.js 13 frontend, FastAPI orchestrator, no external LLM key required.
The Impact
Selected as a Builder submission in the Kapruka Agent Challenge 2026 alongside 700+ developers, delivering real guest checkouts with millisecond client-side re-optimisation when the budget changes.
A multilingual shopping agent built on the Kapruka Model Context Protocol that converts natural-language intent into four budget-aware cart plans and a live guest checkout.
Technical Engine
Sustainable Vision: Multitask Scene & Emission Model
Machine Learning Engineer & Researcher
The Challenge
Carbon intensity cannot be read directly from pixels and has orders of magnitude fewer labels than scene classification, making an independently trained emission model impractical.
The Solution
Architected a multitask network with a Places365-pretrained trunk and three task heads under a weighted loss, then fine-tuned only the emission head on the Intel Image Classification dataset with the trunk frozen to prevent catastrophic forgetting.
The Impact
Produced a deployable checkpoint that delivers scene, attribute, and emission predictions in a single forward pass while preserving full Places365 classification quality.
A single ResNet-50 backbone jointly predicting Places365 scene categories, binary scene attributes, and five-class carbon emission levels from one photograph.