Agentic AI · Generative AI · MLOps Engineer

PRAKHAR
BATWAL

AI-focused CS undergraduate experienced in LLM-powered AI agents, Generative AI applications, agentic workflows, prompt engineering, RAG, structured LLM output, and end-to-end MLOps pipelines with TensorFlow, Scikit-Learn, FastAPI, Docker, and CI/CD.

Get in TouchSee My Work
Problem solving
400+ LeetCode & CodeChef
Hackathons
Top 10 out of 700+ teams (Top 1.4%)
Certifications
3x Azure AI, AWS, MongoDB

Background

I am an AI-focused Computer Science & Engineering undergraduate at Parul University, Vadodara (Aug 2023 – Apr 2027) specializing in Artificial Intelligence, LLM-powered AI agents, and end-to-end MLOps pipelines.

My expertise spans architecting multi-agent workflows, dense semantic retrieval (RAG), structured LLM reasoning, and production-grade machine learning pipelines. Combining classical ML, deep learning, and structured LLM workflows with robust MLOps practices, I build intelligent systems that remain verifiable, safe, and low-latency under real-world scale.

“Software engineering is about writing clean, auditable, maintainable, and self-documenting systems that remain reliable under real-world scale.”

Education

Parul University

Vadodara, India

B.Tech in Computer Science & Engineering (AI Specialization)

Aug 2023 – Apr 2027

Resumes

3 tailored versions — preview and download
Updated Resume (AI Specialization)

AI & MLOps Engineer

Agents, RAG, MLOps, TensorFlow, FastAPI

Tailored for ML roles

Machine Learning & RecSys

Two-Tower models, FAISS, SHAP, Keras

Tailored for backend roles

Software Engineer

FastAPI, PostgreSQL, Docker, Microservices

What I Do

8 focus areas
[01]

Agentic AI & LLMs

Multi-agent orchestration, tool-calling, agentic workflows, MCP, and structured outputs.

LangGraph · LangChain · Gemini · Vertex AI · Nvidia NIM

Benchmark99.8%_RECOVERY
[02]

Hybrid Support Agent

Hybrid AI pipeline combining classical ML, FAISS vector search, and Gemini with strict safety rules.

Python · Scikit-Learn · FAISS · Gemini 2.5 Flash · Pydantic

Accuracy / Macro F190.5% / 91.05%
[03]

Neural Search & RecSys

Two-Tower neural network mapping 200K+ ratings into a 64-D embedding space with FAISS ANN.

TensorFlow · Keras · FAISS · SentenceTransformers · FastAPI

Hit Rate@1025.2%
[04]

Explainable ML

VotingRegressor ensemble combining XGBoost and LightGBM with per-feature SHAP attributions.

Python · XGBoost · LightGBM · SHAP · Scikit-Learn · Docker

R² Score0.847
[05]

NLP & Linguistic Bias

Dual-output credibility and 5-class bias classifier with NLTK VADER and token-level Gemini API.

Python · Scikit-Learn · TF-IDF · NLTK · Gemini API · pytest

RegressionsZERO_DEFECT
[06]

MLOps & CI/CD

Hermetic container builds, automated model tracking, data drift monitoring, and GitHub Actions CI.

MLflow · DagsHub · Docker · GitHub Actions · Cloud Run

DeploymentSUB_50MS
[07]

Dense Semantic Retrieval

Sub-millisecond semantic search over 12K+ dialogue embeddings with 81% Hit@3 and O(log N) ANN.

FAISS · SentenceTransformers · NumPy · Python

Recall81%_HIT@3
[08]

High-Performance APIs

Asynchronous REST APIs with Pydantic v2 validation, SQLAlchemy ORM, and PostgreSQL persistence.

FastAPI · PostgreSQL · Docker · JWT · Python

Latency<10MS_OVERHEAD
Featured Project

Saarthi

End-to-end AI platform for project discovery, feasibility analysis, architecture blueprinting, and runnable codebase generation. Features automated Repo Forge export via GitHub API with transactional Git commits and resilient JSON auto-repair for streaming LLM outputs with 99.8% recovery.

01
Discovery & Feasibility

Natural language requirement ingestion and feasibility scoring.

02
Architecture Blueprint

Interactive system diagrams, API contracts, and schema design.

03
Database Architect

PostgreSQL schema mapping and relational index optimization.

04
Codebase Generation

Full-stack FastAPI backend and React 19 type-safe frontend.

05
Repo Forge & Auto-Repair

GitHub API commit transactions with 99.8% streaming JSON recovery.

React 19TypeScriptFastAPIPythonPostgreSQLDockerNvidia NIMSQLAlchemyPydantic v2
View Code →Live Demo
28AGENTSHow Saarthi's agents work together

Skills & Tools

24 technologies

Agentic AI & LLMs

Multi-Agent Systems95%

LangGraph orchestration

Google Gemini & Vertex AI94%

Structured generation

RAG & Embeddings92%

Dense vector retrieval

Tool-Calling & MCP90%

Agentic workflows

Prompt Engineering92%

Few-shot & guardrails

Nvidia NIM & Azure AI88%

Inference optimization

ML & Data

Scikit-Learn94%

90.5% intent accuracy

TensorFlow & Keras88%

Two-Tower RecSys

FAISS Vector Search92%

81% Hit@3 over 12K+

XGBoost & LightGBM90%

R² = 0.847 ensemble

SHAP Explainability88%

Feature attributions

SentenceTransformers90%

Dense representation

MLOps & Deployment

FastAPI & REST APIs95%

Async sub-50ms inference

Docker & Cloud Run92%

Hermetic containerization

GitHub Actions CI/CD90%

Automated test suites

MLflow & DagsHub88%

Model registry & drift

Pydantic v294%

Type-safe schemas

Linux & Shell88%

System automation

Languages & Tools

Python96%

Core engineering runtime

TypeScript & React 1990%

Type-safe UI systems

SQL (PostgreSQL / MySQL)88%

Relational persistence

MongoDB & Redis86%

Document & cache layers

pytest92%

Zero regression testing

Git & GitHub API92%

Automated repo forge

Projects

7 selected projects
Agentic AI & RAG90.5%_ACC / 81%_HIT@3
  • Architected a hybrid AI agent pipeline combining classical ML, dense semantic retrieval (RAG), safety guardrails, and LLM generation.
  • Trained TF-IDF + Logistic Regression intent classifier achieving 90.5% accuracy and 91.05% Macro F1; built FAISS retrieval over 12K+ dialogues with 81% Hit@3.
  • Integrated Gemini 2.5 Flash with Pydantic structured JSON and negative policy constraints, achieving 4.70/5.00 safety with zero hallucinations; automated 45.5% of inquiries.
PythonScikit-LearnSentenceTransformersFAISSGemini 2.5 FlashPydanticpytest
View Code →
Agentic AI & Full-StackTOP_10 / 700+_TEAMS
  • Built an end-to-end AI platform for project discovery, feasibility analysis, architecture blueprinting, and runnable codebase generation.
  • Engineered FastAPI backend with SQLAlchemy, Pydantic v2, JWT auth, and async REST APIs backed by PostgreSQL; React 19 frontend with type-safe routing.
  • Developed automated Repo Forge export via GitHub API with transactional Git commits; built resilient JSON auto-repair for streaming LLM outputs with 99.8% recovery.
React 19TypeScriptFastAPIPythonPostgreSQLDockerNvidia NIM
View Code →Live Demo →
Deep Learning & RecSysHR@10_25.2% / MRR_0.1078
  • Trained a two-tower neural network on 200K+ ratings using TensorFlow/Keras, projecting users/items into a 64-D embedding space.
  • Achieved Hit Rate@10 of 25.2% and MRR of 0.1078; replaced brute-force with FAISS ANN index cutting retrieval to O(log N).
  • Reduced CPU preprocessing by 95% via optimized vectorization; deployed async FastAPI + Docker with GitHub Actions CI.
PythonTensorFlow/KerasFAISSSentenceTransformersFastAPIDocker
View Code →
Ensemble ML & Explainable AIR²_0.847 / SUB_50MS
  • Achieved R² = 0.847 with VotingRegressor (XGBoost + LightGBM) augmented by geospatial features and KMeans clustering; delivered sub-50 ms inference via Dockerized async FastAPI.
  • Appended per-feature SHAP contributions to every API response for explainability; SQLite persistence for drift monitoring with CI/CD tests.
PythonXGBoostLightGBMSHAPScikit-LearnFastAPIDocker
View Code →
NLP & Generative AIDUAL_OUTPUT_CLF / ZERO_REGRESSION
  • Trained dual-output classifier (credibility + 5-class bias) with TF-IDF and grid-search sweeps, optimizing F1 on imbalanced news corpora.
  • Layered NLTK VADER sentiment and async Gemini API calls for LLM-powered token-level linguistic bias breakdowns alongside confidence scores.
  • Built JS/HTML5 dashboard with bias charts and confidence heatmaps; enforced zero regressions via full pytest suite with modular layout.
PythonScikit-LearnTF-IDFNLTKGemini APIFastAPIpytest
View Code →
Distributed SystemsSUB_50MS_LOOKUP
  • Priority event queues via Kafka with distributed rate limiting.
  • Dead Letter Queue (DLQ) & exponential backoff recovery.
Java 17Spring BootKafkaRedisMySQLDocker
View Code →
MLOps & SecurityLIFECYCLE_DRIFT
  • Automated model retraining loops with data drift alerts.
  • Experiment registry and staging tracking via MLflow.
MLflowDagsHubFastAPIMongoDBDocker
View Code →

Highlights

Certified

Microsoft Certified: Azure AI Apps & Agents Developer Associate

Demonstrated expertise in building AI agents, multi-agent orchestration, Azure OpenAI Service integration, and generative AI application architecture.

Certified

MongoDB Certified Developer

Validated proficiency in building high-performance document data models, transactional schemas, aggregation pipelines, and indexing strategies.

Certified

AWS Certified Cloud Practitioner Essentials

Certified in foundational AWS cloud architecture, compute instances, security controls, containerized workloads, and cloud economics.

2024

Elevate Hackathon (Hack2Skill) — Top 10 / 700+ Teams (Top 1.4%)

Designed and deployed Saarthi, pitching an autonomous AI co-founder platform under competitive hackathon conditions to finish in the top 1.4% of 700+ teams nationwide.

Continuous

400+ Problems Solved on LeetCode & CodeChef

Active competitive programmer with 400+ algorithmic problems solved covering dynamic programming, graph algorithms, trees, and advanced data structures.

“Software engineering is about building systems people can trust.”

Auditability · Scalability · Fault Tolerance · Deterministic Guardrails