AI Systems Online

Hetvi Doshi

> AI Engineer currently

I build intelligent systems that move from prototype to production — engineering RAG pipelines, deploying LLMs, and shipping AI that actually works at scale.

View My Work
Let's Connect →
GitHub LinkedIn Email
~/ai-engineer — zsh
$
MODELS DEPLOYED
12+
UPTIME
99.9%
Scroll
SIH 2025 Finalist
Smart India Hackathon
Top 3000 / 60,000+
Amazon ML Summer School
AI System Builder
AI Engineer
Python TensorFlow Flask PostgreSQL Tailwind CSS OpenAI API LangChain Docker Scikit-Learn Git & CI/CD
ENGINEERING PROOF

My Projects

Not just demos. Production-grade AI systems engineered to solve complex, real-world problems with measurable impact.

SIH 2025 Finalist

AI-Based DPR Evaluation System

[PROBLEM]

Manual Detailed Project Report (DPR) compliance and risk checks took hours and were prone to human error.

[SOLUTION]

Automated hybrid PDF extraction and semantic chunking for scanned, multilingual reports.

[CHALLENGE]

Handling noisy, multilingual scanned documents while maintaining explainable, auditable validation.

[OUTCOME]

Cut review time from hours to minutes using a hybrid RAG + rule-based engine.

Python NLP RAG FastAPI
Architecture Flow
PDF Input Hybrid RAG Rule Engine
Multi-Modal AI

NURA Wellness App

[PROBLEM]

Mental health risk assessment requires synthesizing disparate, complex data signals in real-time.

[SOLUTION]

Integrated behavioral, wearable, Wi-Fi, and conversational signals for automated risk assessment.

[CHALLENGE]

Orchestrating multi-modal data ingestion and maintaining stateful, fault-tolerant LLM reasoning workflows.

[OUTCOME]

Deployed LangGraph workflows combining rule-based analysis with LLM-powered reasoning.

Python FastAPI LangGraph Groq LLM SQLite
Architecture Flow
Sensors LangGraph Risk Score
NL-to-SQL Agent

Slack AI Data Bot

[PROBLEM]

Business teams needed instant data insights without writing complex SQL queries or leaving Slack.

[SOLUTION]

Converted natural language business queries into SQL and delivered insights directly within Slack.

[CHALLENGE]

Ensuring high reliability, preventing SQL injection, and managing LLM hallucination in NL-to-SQL translation.

[OUTCOME]

Implemented caching and fault-tolerant error handling, drastically improving query performance.

Python FastAPI LangChain PostgreSQL
Architecture Flow
Slack API NL-to-SQL Postgres
ENGINEERING METHODOLOGY

How I Build AI Systems

Not just models, but production-ready pipelines. A rigorous, repeatable approach to moving from raw data to scalable intelligence.

01

Discover

Audit data availability, define latency/accuracy metrics, and establish technical feasibility boundaries.

02

Architect

Design the pipeline. Select base models, vector databases, and orchestration frameworks (e.g., LangChain).

03

Build

Develop core logic. Implement RAG retrieval, prompt engineering, and robust, typed API endpoints.

04

Evaluate

Rigorous testing. Measure hallucination rates, token usage, and response latency against baselines.

05

Deploy

Containerize (Docker), orchestrate, and set up CI/CD pipelines for seamless, zero-downtime rollout.

06

Iterate

Monitor data drift, gather user feedback, and continuously fine-tune or optimize inference costs.

0%
Reduction in Manual Data Entry
0.0%
OCR Extraction Accuracy
Top 0
Out of 60,000+ Competitors
0+
Years Building AI Systems
SYSTEM BLUEPRINT

Production Architecture

Not just models, but scalable, fault-tolerant pipelines. Visualize how a request flows from the client to the AI agent and back.

architecture_flow.svg
User Frontend API Gateway AI Agent Vector DB LLM Models Ext. Services
System Status

Idle

Click "Simulate Request" to trace a real-world query through the production AI pipeline, from ingestion to model inference and external tool execution.

Waiting...

Design Principles Applied

  • Fault Tolerance: Graceful degradation if the LLM times out or Vector DB is unreachable.
  • Observability: Every step logged with latency metrics and token usage for continuous optimization.
  • Security: PII redaction at the API Gateway before any data reaches the AI Agent layer.
ENGINEERING CAPABILITIES

Technical Expertise

Organized by core engineering domains. Not just a keyword wall, but the actual systems I design, build, and scale.

LLM & Agentic Systems

Designing context-aware AI agents and retrieval-augmented generation (RAG) pipelines that transform unstructured enterprise data into actionable, structured intelligence.

LangChain RAG Architectures Agent Orchestration Prompt Engineering

AI & Machine Learning

Developing predictive models, NLP pipelines, and multimodal systems for real-world operational efficiency and forecasting.

TensorFlow Scikit-Learn NLP Multimodal ML

Backend & Distributed Systems

Architecting robust, secure, and high-performance API layers to serve AI models at scale with low latency and strict typing.

Python / FastAPI Flask PostgreSQL REST / GraphQL

Data Engineering & Cloud/DevOps

Building automated data pipelines (OCR, ETL) and deploying containerized applications with CI/CD for zero-downtime, scalable infrastructure.

Docker AWS / GCP CI/CD Pipelines Vector DBs (Pinecone) OCR Pipelines
ENGINEERING STACK

Tools & Technologies

The precise, production-grade toolkit used to build, deploy, and scale intelligent systems.

Core Languages

Primary languages for data processing, backend logic, and frontend interactivity.

Python TypeScript JavaScript SQL

AI / Machine Learning

Frameworks and APIs for model training, RAG pipelines, and LLM orchestration.

PyTorch Transformers OpenAI API LangChain Scikit-Learn

Backend & APIs

Robust, typed, and scalable server architectures to serve AI models at low latency.

FastAPI Node.js PostgreSQL Redis

DevOps & Infra

Containerization, CI/CD pipelines, and cloud orchestration for zero-downtime deployments.

Docker AWS GitHub Actions Terraform

Supporting Ecosystem

Git & GitHub • Linux / Bash • Tesseract OCR • Pandas / NumPy • Matplotlib • Jupyter • VS Code • Postman
ENGINEERING_MANIFEST.md

Build for production, not demos.

The core principles that separate hobbyist scripts from scalable, reliable AI systems.

01

Evaluation > Prompts

AI needs rigorous, automated evaluation, not just clever prompting. If you can't measure hallucination rates, token usage, and latency against a baseline, you don't have a product—you have a toy.

02

Measure Before Optimizing

Profile first, optimize second. Bottlenecks in AI pipelines are rarely where you think they are. Guesswork is for amateurs; empirical data drives architectural decisions.

03

Simplicity Before Abstraction

Complexity is technical debt. Start with the simplest pipeline that solves the problem (even if it's a basic RAG setup). Scale and abstract only when the metrics demand it.

04

Automate the Mundane

If an engineering task is repeated more than twice, it belongs in a script, an AI agent, or a CI/CD pipeline. Human brain cycles should be spent on architecture, not manual data cleaning.

05

Architecture Serves the Product

The fanciest model is useless if it takes 10 seconds to respond or costs $5 per query. Reliability, latency, and cost-efficiency are not afterthoughts; they are primary constraints that dictate the architecture from day one.

system_status: operational

"The first version should embarrass you slightly. If it doesn't, you waited too long to launch. Ship, measure, iterate."

INTERACTIVE DEMO

RAG Pipeline Simulator

Don't just take my word for it. Experience how a production Retrieval-Augmented Generation (RAG) system processes a query, retrieves context, and synthesizes an answer in real-time.

rag_pipeline_simulator.js
Idle
2. Embedding text-embedding-3-small
Waiting for query...
3. Vector Retrieval Pinecone (Top-K=3)
doc_policy_v2.txt Sim: 0.00

All PII data is automatically redacted at the API gateway using regex patterns before being passed to the LLM context window.

doc_architecture.md Sim: 0.00

The system utilizes a hybrid search approach combining dense vector embeddings with sparse BM25 keyword matching.

doc_compliance.pdf Sim: 0.00

GDPR and HIPAA compliance requires strict data isolation and encryption at rest for all vector database entries.

DEPLOYMENT HISTORY

Engineering Timeline

A chronological log of technical contributions, system builds, and continuous learning in production AI environments.

Jun 2026 – Present
Surat (Hybrid)

AI Engineer Intern

Active
Samniti Technologies
  • ▹ Contributing to core AI-driven product development and end-to-end automation initiatives.
  • ▹ Scaling machine learning workflows and gaining hands-on industry experience in a hybrid production environment.

AI / ML Intern

Completed
Wizcoder • Surat (On-site)
  • ▹ Built an NLP sales analytics assistant featuring MCP-based dynamic routing across multi-table datasets.
  • ▹ Engineered an NL-to-query pipeline automating structured database queries for real-time business insights.
  • ▹ Designed scalable data architecture enabling conversational business intelligence extraction.
NLP MCP Routing NL-to-SQL
Aug 2025
Remote

ML Mentee

Top 5% (3k/60k+)
Amazon ML Summer School
  • ▹ Selected among 3,000 out of 60,000+ global applicants for an intensive, expert-led machine learning curriculum.
  • ▹ Gained deep, hands-on exposure to advanced ML paradigms: Deep Neural Networks (DNNs), Large Language Models (LLMs), Causal Inference, and Reinforcement Learning.
DNNs LLMs Causal Inference Reinforcement Learning
SYSTEM VALIDATION

Achievements & Credentials

Quantifiable results from competitive environments and industry-verified technical specializations.

PERFORMANCE METRICS

Finalist 2025

Smart India Hackathon (SIH)

MDoNER, Govt. of India • AI-based DPR Evaluation System

Top 0%
National Level
Top 5% 2025

Amazon ML Summer School

Selected for advanced machine learning curriculum and practical applications.

0
Out of 60,000+
Semi-Finalist 2026

ET AI Hackathon

The Economic Times • Competed in high-scale AI problem-solving.

Top 0%
Out of 54,619+ Teams
Ranked 2025

Amazon ML Challenge

Global competitive machine learning challenge.

#0
Out of 6,000+ Teams

VERIFIED CREDENTIALS

OCI Data Science Professional

Oracle

Verified

Data Analytics

IBM SkillsBuild × CSRBOX

Verified

Data Analysis and AI/ML

SAP × Edunet

Verified
TOTAL COMPETITORS OUTRANKED
120,000+
System Uplink

Have a complex
engineering problem?
Let's build the solution.

I specialize in turning ambiguous requirements into scalable, production-ready AI systems. If you need an engineer who understands both the underlying math and the deployment infrastructure, let's talk.

Initiate Contact
Download Resume
Connection Matrix
Email
hetvidoshi2205@gmail.com
LinkedIn
linkedin.com/in/hetvi-doshi22
GitHub
github.com/hetvidoshi22
Location
Surat, Gujarat, India
Mon - Sat: 10:00 AM - 7:00 PM (IST)