Skip to content

Razin Chinwala

Full-stack AI engineer

I build backends and AI agents that keep working when things go wrong.

About

I'm Razin, a full-stack AI engineer in Ahmedabad. I studied computer science with a focus on AI and machine learning, and over the last few years I've moved from training models to shipping the systems around them.

My first real ML work was at IBM SkillsBuild, where a disease-diagnosis model I built reached 86% accuracy and our team placed in the top 6 of 650+ across India. At BISAG-N I worked with satellite imagery, comparing YOLO and U-Net for rooftop detection. U-Net won with an F1 score of 0.89.

Today at Code Colonies I build the backend side of AI products: a multi-tenant WhatsApp automation SaaS, an autonomous sales agent, and the gateway that rate-limits and meters every LLM call. The part I enjoy most is making these systems correct under failure, so a crash never sends a message twice or spends the same budget again.

By the numbers

database migrations I designed for WhatsAuto
98 of 113
routes moved off legacy Express in under a week, with zero regressions
68+
of 650+ teams at IBM SkillsBuild
Top 6
F1 score on satellite rooftop detection
0.89

What I build

  • AI agents with a human in the loop

    Agents that read email, qualify leads, book meetings and draft proposals, with approval gates before anything irreversible happens.

  • Retrieval and search

    RAG pipelines with hierarchical chunking and reranked, cited answers, plus image search.

  • Multi-tenant SaaS backends

    Tenant isolation with PostgreSQL row-level security, role-based access, webhook ingestion and idempotent outbound queues.

  • Workflows that survive crashes

    Durable Temporal workflows, two-phase reservations with row locks, and crash-recovery reconciliation.

Experience

CCCurrent role

Code Colonies

Ahmedabad

Associate Fullstack AI Consultant

Apr 2026 to now

Leading AI-heavy backend work: a WhatsApp automation SaaS, autonomous sales agents, and the infrastructure that keeps LLM workflows safe and metered.

  • Led development of WhatsAuto, a multi-tenant WhatsApp automation SaaS: chose the stack, designed the database schema (98 of 113 migrations), and deployed it on a Contabo VPS.
  • Built tenant-isolated RBAC with PostgreSQL row-level security, plus a Meta WhatsApp Cloud API integration with webhook ingestion, an idempotent outbound queue, and template sync.
  • Built an AI auto-reply agent that answers from each tenant's knowledge base with human takeover, plus a visual flow builder for message automation and campaign scheduling.
  • Migrated an internal sales CRM from legacy Express.js (68+ routes) to Fastify, Prisma, and PostgreSQL with JWT RBAC in under a week, with zero regressions.
  • Built an autonomous AI sales agent that reads and replies to email, qualifies leads, books meetings, and drafts proposals, with human approval gates.
  • Designed a multi-LLM gateway with distributed rate limiting, priority lanes, and token metering, used by Temporal lead-discovery and campaign workflows.
  • Built a distributed multi-agent scheduling engine for a community platform in 1 day for a stakeholder demo, with probabilistic, activity-aware scheduling.
  • Implemented two-phase action reservations with row locks and crash-recovery reconciliation to prevent duplicate AI actions and budget leaks.
  • Built search-by-image for an e-commerce platform using AWS Bedrock embeddings, a pgvector HNSW index, and an SSRF-hardened image proxy.
  • Built an event-driven "frequently bought together" engine (Jaccard similarity to reduce popularity bias), plus order, shipping, and tracking notifications through WhatsAuto.
  • TypeScript
  • NestJS
  • Fastify
  • Prisma
  • PostgreSQL
  • Temporal.io
  • Redis
  • React
  • Docker

Internships

BN

BISAG-N

Gandhinagar

Research Intern: Computer Vision

Feb–May 2025 (4 months)

  • Designed a rooftop detection pipeline on satellite imagery comparing YOLOv11 (S/M), YOLOv8 (S/L), and U-Net (ResNet-152). U-Net reached the best F1 of 0.89 vs 0.73–0.76 for YOLO.
  • Improved feature extraction with classical and deep-learning filters, and regularized predictions with minimum bounding rectangles from model and HED edge outputs, refined by geospatial queries.
  • PyTorch
  • QGIS
  • NumPy
  • Rasterio
  • Scikit-Learn
  • Matplotlib
IBM

IBM SkillsBuild

Remote

AI Intern

Jun–Jul 2024 (2 months)

  • Developed a disease diagnosis model with IBM AutoAI on 257K+ rows, 371 features, and 757 diseases, reaching 86% accuracy.
  • Integrated IBM watsonx Assistant for conversational specialist recommendations, and placed in the winning top 6 of 650+ teams across India.
  • IBM watsonx AI
  • IBM Cloud
  • AutoAI
  • Data Preprocessing

Projects

2

Agentic AI Chatbot with RAG

2026

Full-stack web app

A production-grade agentic chatbot with a 4-path state machine, hierarchical RAG with verified citations, from-scratch auth, and a human-in-the-loop Jira workflow. Piloted live with 8 users.

  • Hybrid router (regex fast path + structured classification) across chat, RAG, weather, and bug triage, cutting redundant LLM calls. Streams tokens over SSE with disconnect safety.
  • Dual Redis/PostgreSQL memory with token-budget compaction and rolling sliding-window summarization.
  • Conversation-scoped RAG: hierarchical chunking, local ONNX BGE embeddings on CPU, Pinecone multi-tenant namespaces, and Cohere Rerank v3 for ranked, verified citations.
  • Auth from scratch: OAuth2/JWT, refresh-token rotation, SHA-256 hashed token storage, multi-session device pruning, and admin RBAC.
  • Human-in-the-loop chat → Jira workflow with row-level locks and atomic Redis rate limiting to prevent duplicate tickets. Voice I/O via the Web Speech API.
  • Multi-stage Docker images built with uv, deployed on AWS (EC2, RDS PostgreSQL, S3, CloudFront).
  • Next.js
  • FastAPI
  • LangGraph
  • LangChain
  • Redis
  • PostgreSQL
  • Pinecone
  • Cohere
  • Hugging Face
  • Docker
  • AWS

Sentiment Analysis of Movie Reviews

2024

NLP research

Fine-tuned transformer models for binary sentiment classification on the IMDB dataset.

  • Fine-tuned BERT and RoBERTa on IMDB, reaching 93.86% and 94.84% accuracy.
  • Python
  • PyTorch
  • Transformers

Stack

Languages

  • TypeScript
  • JavaScript
  • Python
  • SQL
  • HTML
  • CSS

Backend and frontend

  • Node.js
  • NestJS
  • Fastify
  • FastAPI
  • Express.js
  • Prisma
  • TypeORM
  • Temporal.io
  • BullMQ
  • React
  • Next.js

AI and LLM

  • LangChain
  • LangGraph
  • RAG
  • AI Agents
  • OpenAI
  • Gemini
  • Mistral
  • AWS Bedrock
  • pgvector
  • Pinecone
  • Hugging Face
  • PyTorch

Data and DevOps

  • PostgreSQL
  • Redis
  • Docker
  • AWS
  • Railway
  • Vercel
  • Git

Education

AIIE

Adani Institute of Infrastructure Engineering

B.E. Computer Science & Engineering (AI & ML)

2021–2025 CGPA 8.94 / 10

IJSC

Infocity Junior Science College

Higher Secondary (XII)

2019–2021 86%

Certifications

  • Introduction to Node.js, v3
  • TypeScript: From First Steps to Professional
  • Professional CSS: Build a Website from Scratch

Contact

Building something with LLMs that has to work in production?

I'm happy to talk through agents, retrieval, or the backend that holds them together. Email is the fastest way to reach me.

Email me
Résumé