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.
| Building RAG platforms with solid ingestion, retrieval, and grounding so answers stay traceable to source content. | Orchestrating tool using agents with LangGraph, MCP, and PydanticAI for workflows that need planning and control. |
| Provider failover, circuit breaking, caching, rate limiting, cost aware model routing, and dashboards for LLM traffic in production. | FastAPI backends with React and Next.js frontends, backed by PostgreSQL and Redis, shipped in containers. |
Languages
AI and LLM Frameworks
Backend and Frontend
Data and Storage
Machine Learning and MLOps
Cloud, DevOps and Observability
|
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.
|
