Open to opportunities · Jersey City, NJ

Bhakti Shelke Senior AI Engineer

5+ years building production-grade LLM systems, RAG architectures, and agentic AI pipelines. Currently at HCZ, New York — cutting hallucinations by 35%, analyst workload by 60%, and model iteration cycles from 2 days to under 4 hours.

View My Work LinkedIn Medium

Years Python

35%

Hallucination Cut

60%

Research Time Saved

3.60

MS GPA · Pace

// about me

Building AI that
ships and scales

I'm a Senior Python Developer and AI Engineer with 5+ years delivering scalable data ingestion pipelines, LLM-powered classification systems, and production-grade RAG architectures.

At Harlem Children's Zone in New York, I architect agentic AI systems using LangGraph, AWS Bedrock, and Llama — building orchestration layers, vector search pipelines, and microservices that connect LLMs to enterprise systems at real scale.

Previously at ZiSystech (FinTech) and JSW Steel. MS in Data Science, Pace University (GPA 3.60). BE in Computer Science, University of Mumbai (GPA 3.80).

inlinkedin.com/in/bhakti-shelke Mmedium.com/@bhaktishelke2024 tel+1 551-430-9955

Python & Backend

Python (5+ yrs)FastAPIREST APIsMicroservicesDockerCI/CDGitHub Actions

LLMs & AI

LangChainLangGraphAWS BedrockLlamaHugging FacePrompt EngineeringLLMOps

RAG & Vectors

PineconeFAISSChromaEmbeddingsRAG ArchitecturesEval Frameworks

Big Data & Cloud

PySparkHadoopHiveDelta LakeSnowflakeAWS S3/LambdaPostgreSQL

Languages

PythonSQLPySparkJavaScriptC++R

// selected work

Projects

Production systems and independent builds — real numbers attached.

🤖

↗

DeepFetch AI — Multi-Agent LLM Product Intelligence Engine

Independent multi-agent Python platform using LLMs, embeddings, and vector search to automate large-scale e-commerce feature extraction. 5-stage agent reasoning loop with full LLMOps evaluation pipeline for production monitoring.

88%+ extraction accuracy vs ~60% manual baseline

45% reduction in post-processing correction rate

FAISS + Chroma RAG retrieval layer for unstructured data

PythonLLMsFAISSChromaLangChainLLMOpsMicroservices

# agent reasoning pipeline
Stage 1 → Data Ingestion
Stage 2 → Feature Extraction
Stage 3 → Prompt Refinement
Stage 4 → RAG Retrieval
Stage 5 → LLMOps Evaluation
accuracy: 88%+
correction rate: ↓45%

🔍

↗

Modular RAG Pipeline

Pinecone/FAISS vector search pipeline at HCZ that cut analyst manual research time by 40–60%, eliminating a full sprint of manual work per project cycle.

40–60% reduction in manual research time

Pinecone + FAISS hybrid retrieval

PineconeFAISSPythonLangChainAWS Bedrock

⚙️

↗

LangGraph Agentic Orchestration Layer

State-machine orchestration for deterministic multi-step agent execution with automatic error recovery — reducing failure-driven reruns by ~50% in production.

~50% fewer agent failure reruns

Deterministic multi-step execution

LangGraphPythonAWS LambdaLlama

📊

↗

LLMOps Prompt Evaluation Harness

Reusable evaluation framework with rubric scoring and retrieval recall metrics that slashed model iteration from 2 days to under 4 hours. Adopted across 3 internal AI products.

2 days to under 4 hours per iteration

Adopted across 3 internal products

PythonLLMOpsRubric ScoringEval Frameworks

⚡

↗

PySpark ETL / ELT Pipeline

ETL/ELT pipelines across 4 source systems using PySpark and SQL at ZiSystech FinTech, reducing query latency by 40% and enabling real-time reporting.

40% query latency reduction

4 source systems integrated

PySparkSQLHiveDelta LakeSnowflake

// career

Experience

Where I've built, shipped, and driven measurable impact.

Harlem Children's Zone (HCZ)

AI Engineer / Senior Python Developer · New York, NY

Apr 2024 – Present

Architected Python-based data ingestion pipelines integrating structured and unstructured data into LLM workflows, enabling scalable classification and summarization across enterprise datasets.
Implemented Llama via AWS Bedrock for text classification and summarization — reduced hallucination rate by 35%+ through rubric scoring and retrieval recall evaluation frameworks.
Built modular RAG pipelines (Pinecone/FAISS) cutting analyst manual research time by 40–60%, eliminating a full sprint of manual work per project cycle.
Engineered LangGraph state-machine orchestration for deterministic multi-step agent execution with automatic error recovery — reduced failure-driven reruns by ~50%.
Shipped Python microservices and REST APIs connecting LLM agents to 5+ enterprise systems (AWS Lambda, S3, Bedrock, Snowflake) with zero-downtime, auditable handoffs.
Built reusable prompt evaluation harness (LLMOps) cutting model iteration from 2 days to under 4 hours; adopted across 3 internal AI products.

ZiSystech Pvt Ltd (FinTech)

Senior Software Engineer · India

Jun 2021 – Jan 2023

Built ML-integrated Python pipelines automating 3 manual data quality workflows and eliminating ~15 hours/week of analyst review.
Designed ETL/ELT pipelines across 4 source systems using PySpark and SQL — reduced query latency by 40% and enabled real-time reporting.
Developed RESTful microservices with zero post-release critical bugs.
Built model evaluation dashboards reducing review cycles by 2 rounds per release.

JSW Steel Pvt Ltd

Software Developer · India

Jan 2018 – Jan 2021

Refactored legacy C++ backend bottlenecks achieving 30–35% reduction in system response time.
Built 3 JavaScript UI modules adopted by 2 product teams; raised test coverage from ~10% to 65%.

Bhakti Shelke Senior AI Engineer

Building AI that
ships and scales

Projects

Experience

On Medium

Education

Let's build something
remarkable together.

Bhakti Shelke Senior AI Engineer

Building AI thatships and scales

Projects

Experience

On Medium

Education

Let's build somethingremarkable together.

Building AI that
ships and scales

Let's build something
remarkable together.