Agentic audience workflows
Built a real-time agentic workflow that turns trader queries and first-party data into custom audiences with R-F-D scores. Integrated data APIs and tools to give the agent relevant, high-quality context.
Bengaluru, India12.97° N / 77.59° E
LokranjanP.
Data Scientist I / MiQ Digital
Machine learning. Agentic AI. Real-world systems.I build production ML and agentic AI systems. From semantic retrieval to intelligent workflows, I like taking an idea all the way from experiment to something people can use.
Experience, projects, skills & education. Ready for your next prompt.
PROJECT WORKFLOWSFIG. 001
garmin.output01 / CONNECT
› garmin_resume_session()
Sample session ready. Garmin data tools available.
AUTO CYCLE01 / 04
01ABOUT
Curiosity, followed
by implementation.
Turning data into systems that do something useful.
At MiQ Digital, I work at the intersection of machine learning, generative AI and MLOps. I build the pipelines and workflows that turn complex data into usable products.
My work spans agentic audience systems, creative intelligence, semantic retrieval, and model optimization. I care about what happens after the notebook: reliable deployment, relevant context, and sensible compute costs.
Outside work, that same curiosity shows up in projects like a Garmin MCP server and an agent learning to play Asteroids. Different problems, the same habit of building to understand.
A little more about my work02EXPERIENCE
From experiment to production.
Building ML and hybrid GenAI systems for data-driven products, with an emphasis on useful intelligence and efficient infrastructure.
MLOps, with less overhead
Productionized automated data refresh and onboarding pipelines. Used Amazon Bedrock Batch Inference to reduce monthly LLM token costs by 70% and remove the LangChain dependency from the pipeline.
From creatives to audiences
Orchestrated an image/video-to-metadata pipeline that normalizes ad creatives and maps them to IAB categories, cutting traders’ manual audience selection time by 50%.
Smaller models, faster relevance
Fine-tuned a MiniLM cross-encoder with soft labels from a RoBERTa model, evaluated it with NDCG and MSE, and packaged it in MLflow. Reduced computation time by 40%.
70%
lower LLM token costs
Bedrock batch inference
50%
less manual selection time
Creative-to-audience pipeline
40%
less computation time
Cross-encoder fine-tuning
03SELECTED PROJECTS
Built to answer “what if?”
Garmin MCP
View Garmin MCP on GitHubA personal fitness intelligence layer. Connect Garmin data to AI tools, understand the patterns, and get useful insights delivered to your inbox.
Under the hood
- Unified real-time and historical Garmin Connect data through Garth’s SSO access layer; adopted by 2,000+ users.
- 15+ analytical tools for physiological and activity metrics, statistical analysis, rolling trends, and visualizations.
- Automated daily, weekly, and monthly insight reports delivered through SMTP.
AsteroidsAI
View AsteroidsAI on GitHubTeaching an agent to survive an arcade universe. An exploration of reinforcement learning, attention, and the space between human and machine gameplay.
Under the hood
- Upgraded a baseline Deep Q-Network to a Transformer-Q Network with attention for multi-object tracking.
- Trained on AWS SageMaker with custom reward–penalty logic for stable Q-learning.
- Recognized as Best Minor Project at NIE in 2024, across 240 teams.
A SMALL INTERNET DETOUR / 03
OnlyMemes
A Flask app for discovering random Reddit memes across multiple subreddits. Redis caching keeps retrieval fast and the scrolling seamless.
04TOOLS & TECHNOLOGIES
The things I reach for.
/ 01
Programming
/ 02
Machine learning
/ 03
AI & search
/ 04
Cloud & infrastructure
05EDUCATION & RECOGNITION
A foundation. A few milestones.
DEC 2021 — SEP 2025
The National Institute of Engineering
B.E. in Information Science
Mysuru, KarnatakaCGPA8.6 / 10
Best Individual Contributor
MiQ Digital · Q4 2025 & Q2 2026
Recognized for technical enhancements and product impact.
Best Minor Project
The National Institute of Engineering
Awarded across 240 teams for the AsteroidsAI project.
A little recognition for work I enjoyed doing.
06LET’S CONNECT
Good ideas start
with a conversation.
Working on something with ML or AI? Want to compare notes, collaborate, or just say hello? I’m always up for a thoughtful conversation.
lokranjandev@gmail.com