I am a Full Stack AI Engineer at Estuate, Inc., based in Bengaluru, India. I design and ship AI systems end to end, from the data and retrieval layer to the API and the user interface. My work centers on production grade Retrieval Augmented Generation platforms, multi agent workflows, and the operational side of LLM applications: reliability, observability, evaluation, and cost control.
I care about building AI features that behave well outside a demo. That means clear architecture, tested failure paths, measurable quality, and systems that a team can run and maintain.
Programming Languages
Generative AI, LLMs and Agentic AI
RAG, Search and Knowledge Systems
LLM Providers and AI Platforms
Machine Learning and Deep Learning
LLM Evaluation, Guardrails and Optimization
Backend, Frontend and APIs
Databases, Caching and Messaging
MLOps, LLMOps and Deployment
Observability, Testing and Reliability
|
Routes each LLM query to the cheapest model tier that answers it acceptably. A fine tuned Laya encoder predicts the minimum sufficient tier (local small, mid tier, or frontier) with calibrated confidence, and low confidence queries escalate to a stronger tier. Labels come from an LLM judge comparing every tier against the frontier answer, and the routing threshold is fixed on a separate calibration split before final testing. On a 60 query held out test split from a 300 query dataset, the router reached 88.3% tier accuracy, retained 91% of frontier quality at roughly 3.6 times lower cost, and temperature scaling reduced calibration error from 0.142 to 0.031. Router latency measured 124 ms at p50 on CPU.
|
|
|
An OpenAI compatible gateway that routes chat completions across multiple LLM providers with health aware failover, hedged requests, deferrable retries, circuit breaking, semantic caching, tenant level rate limits, and full observability.
|
An end to end credit card fraud detection pipeline with data validation, time ordered splits, baseline and XGBoost candidates, MLflow tracking, Azure Machine Learning training, and a PR-AUC quality gate automated through GitHub Actions.
|
|
Hands on exploration of LLM fundamentals, covering tokenization, GPT-2 internals, and text generation.
|
Video generation experiments with ComfyUI on AMD ROCm, documented in notebooks.
|
|
A chatbot built with a Keras Sequential model and a pretrained model.
|
A full run of the Complete Python Pro Bootcamp, building Python fundamentals through daily projects.
|
