01
Observe Before You Optimize
I treat instrumentation, metrics, and debuggability as part of the product surface so systems can be understood before they are tuned.
AI/ML Systems Portfolio
Data to Decisions
Saransh Kumar
Data & ML Systems Engineer
I build production-grade data, ML, and AI systems, from large-scale analytics platforms to intelligent interfaces. My work turns messy data into reliable tools and decisions.
Data Platforms
500+ TB
ETL and analytics pipelines engineered for SmartTV-scale telemetry.
Cost Efficiency
$200K/mo
Cloud spend reduced through serverless Spark and platform modernization.
Decision Reach
100M+ users
SmartTV ad delivery and analytics systems operated across international markets.
About
Engineering machine learning like an operating system, not a one-off model.
My work centers on turning analytical prototypes into production systems that are observable, efficient, and grounded in business constraints. Across large-scale data platforms, ML systems, and intelligent interfaces, I have focused on making machine learning practical enough to ship and reliable enough to trust. I am currently pursuing a Master’s in Data Science at the University of Maryland, College Park, with a focus on machine learning and large-scale data systems.
Design Philosophy
A systems-oriented approach to shipping reliable intelligence.
Every project is framed around observability, interfaces, operator trust, and measurable outcomes.
02
Model the Interfaces
The most durable ML systems are built around clean handoffs between data, services, and operators rather than around a single clever model.
03
Design for Constraints
Latency, cost, governance, and operational trust shape architecture as much as accuracy, especially when the system needs to survive contact with production.
04
Ship Incremental Leverage
I prefer systems that create compounding value: a cleaner pipeline, a better evaluation loop, or a platform primitive that unlocks faster iteration later.
Featured Work
Three case studies on the systems I'm building now.
These stories focus on system framing, architecture, critical flows, and the tradeoffs that shaped each build.
Inspectable Analytics Copilot
Planera
I built Planera as an analytics workspace where planning, SQL generation, validation, and execution stay visible enough for a human to trust the answer.
NL-to-SQLValidation PipelinesLLM OrchestrationExecution Tracing
Trust Surface
Traceable query path
Risk Control
Validation before answer
Sprint Planning Copilot
Contour
I built Contour as a planning copilot that turns vague sprint goals into scoped work, priority signals, and assignment recommendations grounded in how a team actually operates.
Planning SystemsTask DecompositionPrioritizationAssignment Logic
Planning Output
Goal-to-work conversion
Assignment Logic
Skill and capacity aware
Model-Agnostic Adaptive Inference
Prompt-Adaptive Diffusion Optimizer
I built a model-agnostic adaptive inference framework that cuts Stable Diffusion v1.5 latency without training or fine-tuning. It combines local LLM prompt complexity estimation with latent-convergence early stopping, reducing average runtime from 70.84s to 39.17s while preserving CLIP alignment.
PythonPyTorchHugging Face DiffusersOllamaCLIPStreamlit
Latency Reduction
70.84s -> 39.17s
CLIP Alignment
0.6556 adaptive vs 0.6565 baseline
Selected work across data platforms, intelligent interfaces, and applied AI systems.