Open to data science & AI roles
Data Scientist working at the intersection of analytics and applied AI — designing systems that combine embeddings, vector search, LLMs, and multi-agent reasoning to turn messy real-world data into trustworthy decisions.
I'm a data scientist focused on applied AI — the part of the field where embeddings, LLMs, and agents actually have to work on noisy, inconsistent, real-world data. Most of my recent work has been building entity-resolution and reasoning systems at the million-record scale: vector search to find candidates, clustering to group them, and LLMs (with a healthy dose of agentic validation) to make the final call.
Underneath that is a strong analytics foundation — SQL pipelines, Tableau and Power BI dashboards, statistical modeling, classic ML. I think the best AI work comes from people who know what the data actually looks like before they hand it to a model, and I write about both sides of that craft on Medium.
LLM-as-Judge pipelines, multi-agent workflows (Generator / Critic / Improver / Memory), Cohere & Command-family models, RAG with vector DBs.
Semantic matching, multilingual embeddings, DBSCAN clustering with token blocking, scalable similarity search at 1M+ records.
Classification, regression, segmentation; SQL & ETL pipelines; Tableau and Power BI dashboards that decision-makers actually use.
Extended a 1M+ record entity-matching pipeline with a multi-agent validation framework — Generator, Critic, Improver, and Memory agents collaborate to evaluate candidate matches with self-consistency, symmetry checks, and confidence-based routing. Improved accuracy on hard edge cases and cluster-level consistency that single-pass LLM judging missed.
Built a Company 360 system to resolve duplicates and name
variations across 1M+ company records in multilingual sources.
Combines Cohere embed-multilingual-v4 for semantic
representation, DBSCAN with token blocking for candidate
generation, and Cohere Command-A as an LLM verifier — delivering
high-precision matching with interpretable validation.
Supervised ML workflows — classification and regression with scikit-learn — covering preprocessing, model selection, and visualization across multiple datasets to demonstrate practical applications and trade-offs between approaches.
Python-driven customer segmentation and behavioral analysis — Pandas for transformation, scikit-learn for clustering, and Matplotlib for visuals. Surfaced actionable cohorts for retention strategy.
Interactive dashboard providing a comprehensive view of leads, accounts, and opportunities — surfacing high-potential leads by source, industry, and region to lift conversion rates.
End-to-end campaign analytics — impressions, clicks, conversion rates across formats and advertisers — built to drive optimization decisions in near real-time.
Real-time Tableau dashboard backed by complex SQL queries — surfacing job-title KPIs across countries and regions for workforce and market analysis.
.jpg)
Executive-level view of retention, satisfaction, and loss ratio — claims patterns broken down by state and demographic to inform strategic decisions.
Excel-based forecasting tool with VBA automation — statistical models for revenue projection and budget analysis with live data connectivity.
A walkthrough of moving from single-pass LLM validation to a four-agent system (Generator, Critic, Improver, Memory) — and what that buys you on edge cases and cluster consistency.
Read articleHow embeddings, DBSCAN, and Cohere Command-A combine to deduplicate 1M+ multilingual company records — and why the hybrid embedding + LLM approach beat either technique alone.
Read articleAn accessible introduction to ML concepts, algorithm families, and supervised learning — written for readers building their first mental model of the field.
Read articleWhether you need an LLM pipeline that actually holds up in production, a dashboard that gets used, or a fresh pair of eyes on a messy dataset — I'd love to hear about it.
vdevasena1997@gmail.com