Pranav Marla

AI Engineer / Builder

I build AI products that can scale to millions of users.

I currently work as an AI Engineering Consultant at DAZN, a globally leading sports streaming platform. My work is focussed on working on AI pipelines that are highly scalable and robust.

I'm also building a visual reasoning agent that can reason about images and videos using hierarchical reasoning called Kalpa.

Work Experience

DAZN

AI Engineer

Remote, United States Mar 2025 - Present

Worked across DAZN's production AI/ML systems, with primary focus on Agentic AI video products and ML platform infrastructure.
Drove end-to-end implementation across an Agentic AI sports-video compilation platform, building platform infrastructure to support 200k+ users.
Fine-tuned smaller task-specialized LLMs with LoRA using supervised fine-tuning (SFT) on agent interaction data.
Deployed fine-tuned models on AWS Sagemaker, improving agent-task success rate by 5–15% while reducing inference latency by ~40% and costs by ~60% versus larger baseline models.
Built abstraction layers on top of foundation models via AWS Bedrock to standardize prompting, tool use, and response schemas across services, enabling modular extensibility.
Developed an automated testing & monitoring framework for Agentic AI pipelines, reducing manual QA effort by 50%.
Led migration for various user analytics models to AWS Sagemaker, including churn and recommendation workflows, maintaining production batch inference pipelines with Airflow orchestration, monitoring, and cost controls.

Feuji

Generative AI Engineer

Dallas, United States Apr 2024 - Dec 2024

Built and scaled an AI Agent system for IVR menu handling using GPT-4o, leveraging function calling, structured outputs, and real-time LLM inference pipelines to handle 1000+ concurrent requests.
Led development on internal GenAI tooling for the recruitment team, designing structured prompt interfaces, tool invocation patterns, and context handling aligned with Model Context Protocol (MCP) principles, reducing lead processing time by 30%.

Jama Wealth

AI/ML Engineer

Hyderabad, India Jul 2020 - Sep 2022

Designed and automated an end-to-end ML pipeline using Python, Django, SQL, and AWS (EC2 + Lambda), enabling real-time price tracking and execution of hourly equity trading strategies.
Integrated the pipeline with cron-based retraining and batch prediction jobs, improving signal freshness and execution latency by ~40%.
Trained Bi-RNNs, HMMs, and regression ensembles on market data to identify trading patterns and refine signals.
Deployed a production-ready ElasticNet model for real-time trading, boosting backtested returns by ~15% while maintaining low-latency execution and operational reliability.
Built interactive stock chart visualizations with backtesting and performance stats using Django, React.js, TypeScript, and AWS; launched during COVID and increased site traffic by ~70%.
Developed a dynamic method to calculate fund NAVs, reducing computation time to 1/10th of previous version.

Fission Labs

Intern

Hyderabad, India Jun 2019 - Jul 2019

Developed Tweetsock, an app to deduce political alignment from a Twitter user's tweet history using RNNs, deployed using Django, React.js, and AWS.

Education

Northeastern University

Master of Professional Studies in Analytics

Boston, United States Sep 2022 - Dec 2023

Concentration: Applied Machine Learning

Chaitanya Bharathi Institute of Technology

Bachelor of Engineering in Information Technology

Hyderabad, India Aug 2016 - Apr 2020

Projects

Kalpa post

1. Kalpa kalpa.chat

A general LLM agent designed around Socratic reasoning, recursive task decomposition, transparent execution, and self-healing behavior.

The central idea is that complex tasks should not be handled as one-shot completions. They should be unpacked, inspected, challenged, and refined while the user can still see what is happening.

Debate Tree post

2. Debate Tree

marlaman/DebateTree

A framework to steelman, structure, and visualize debates so reasoning quality becomes inspectable instead of rhetorical.

It came from a simple frustration: most arguments optimize for winning rather than for understanding. Debate Tree was an attempt to make disagreement legible.

treesofARC post

3. treesofARC

marlaman/treesofARC

A recursive-agent approach to the ARC Challenge with an emphasis on visualized reasoning and interpretable decomposition.

The project explored how recursive structure could help models treat hard symbolic tasks more like a solvable process than a blind guess.

Side Quests

soundcloud.com/pranav-marla Tracks / mixes

Curated playlist My favourite Electronic music acts of the last decade

Outside work

Music

I am passionate about making and curating music and have been working on music production for the past 5 years. I enjoy working on house music with ecletic influence from different genres and cultures like Classical Carnatic, Jazz and Indie Rock.

Just like engineering, music is a craft that invovles problems solving, and it helps me think in different perspectives.

Outside work

Birdwatching

I recently fell in love with birdwatching. I'm fascinated by the diversity of birds and their habitats. In the world of birdwatching, spark birds are the ones that make you start paying attention to the avian kingdom.

The great blue heron was my spark bird, and it's a bird that still takes my breath away.

Talk to an AI version of me

Ask anything about my work at DAZN, the ideas behind Kalpa, how I think about agents, or what I care about outside work.