Languages
- Python
- JavaScript
- TypeScript
- SQL
- Go
Open to remote roles
Dipu KumarSharma
Software EngineerPython DeveloperAI Engineer
Building scalable backend systems and intelligent AI-powered applications.
- Location
- India
- sdipu231@gmail.com
- Experience
- 6+ years
rag-pipeline · live
- Python
- FastAPI
- Django
- RAG
- LangChain
- Qdrant
- PostgreSQL
- Nuxt 3
- Vue.js
- React
- Node.js
- Go
- Docker
- AWS Lambda
- Celery
- RabbitMQ
- Redis
- MongoDB
- Ollama
- Agentic AI
- GCP
- WebSockets
- Chroma DB
- TypeScript
- 0+
- Years Experience
- 0
- Projects Delivered
- 0+
- Technologies
- 0+
- APIs Built
- 0
- AI & RAG Systems
since Nov 2019
across 5 companies
languages to vector stores
REST, GraphQL, WebSocket
shipped to production
About
Six years of shipping the part users never see
I build the systems that sit behind the product — the APIs, the data pipelines, and increasingly the AI that reasons over them.
Over the past six years I have shipped production software across payments, document intelligence, construction, HR analytics and pharma logistics. My core is Python — FastAPI and Django — paired with Node.js, and lately a lot of applied AI: retrieval-augmented generation, LLM integration and agentic workflows that actually hold up outside a demo.
I care most about the unglamorous parts of engineering. Schema design that will not need a migration in six months. Idempotent webhooks. Retrieval that returns the right chunk instead of a plausible one. Guardrails so a chatbot answers "I do not have that record" rather than inventing it.
I have also led a web team, which taught me that architecture decisions are mostly communication decisions. I am comfortable owning a feature end to end, from the Postgres index up through the Nuxt component that renders it.
Scalable backend architecture
Service-oriented FastAPI and Django systems with async task queues, message brokers and clean separation between transport and domain logic.
Production AI, not demos
RAG pipelines with document parsing, chunking and vector retrieval — built so the model stays inside the provided context and refuses to guess.
Data modelling under pressure
Postgres window and aggregation functions, duplicate-safe ingestion, and Mongo schemas shaped around real query patterns rather than convenience.
Cloud and delivery
AWS Lambda, S3, DynamoDB and GCP deployments, containerised with Docker and shipped through repeatable CI/CD.
Education
B.E. — Computer Science & Engineering
Rajiv Gandhi Prodyogiki Vishwavidyalaya · Bhopal, MP
2019
Higher Secondary — PCM
Uttar Pradesh Board · Ballia, UP
2013
Certifications
- Data Visualization and Machine LearningTutorials Point
- FastAPIUdemy
- Nest.jsTutorials Point
Stack
The tools, and how deep the water goes
Grouped by what they're for rather than by hype. The meters are an honest self-assessment: five blocks means I've run it in production under load, three means I'm productive and still learning the edges.
Backend
- FastAPI
- Django
- Express.js
- Nest.js
- AWS Chalice
AI & Agents
- RAG
- Generative AI
- Agentic AI
- LLM Integration
- LangChain
- Ollama
- LangGraph
Frontend
- Nuxt 3
- Vue.js
- React.js
- Tailwind CSS
Databases & Vector Stores
- PostgreSQL
- MongoDB
- Redis
- Qdrant
- Chroma DB
- MySQL
- DynamoDB
- Supabase
DevOps & Cloud
- Docker
- AWS
- GCP
- Git
- CI/CD
- Azure
- Jenkins
Messaging & Async
- Celery
- RabbitMQ
- WebSockets
Tools & Platforms
- GitHub
- Odoo
- Web Scraping
- LangSmith
Selected work
Fifteen systems that shipped
Client and product work from the last six years — document AI, payment rails, approval engines, high-throughput parsers. These are private codebases, so what you get here is an honest account of what I built rather than a repo link.
filter
01
DEVtrust 2025 — Present
E-Abhilekh
Departmental document management with an AI smart-search layer that answers questions from uploaded PDFs.
- Django
- Python
- Generative AI
- LLMs
- +3
Private client codebase
02
DEVtrust 2025 — Present
Updesh Chatbot
A RAG chatbot engineered to stay inside its context — it says "no record" instead of inventing one.
- FastAPI
- Python
- Nuxt.js
- Chroma DB
- +3
Private client codebase
03
Dreamztech Solutions 2021 — 2024
Flux Pay
Payment gateway backend covering admin, merchant and user modules with multi-way fund collection.
- FastAPI
- Python
- PostgreSQL
- RabbitMQ
Private client codebase
04
UniteCloud 2024 — 2025
RR Parkon
Construction project management system with dependency-aware milestone tracking. Led the web team.
- Express.js
- Node.js
- Nuxt.js
- MongoDB
Private client codebase
05
UniteCloud 2024 — 2025
HRMS Attrition Analytics
Full-stack HRMS that analyses employee attrition and predicts leave, reading live from Odoo.
- FastAPI
- Python
- React.js
- Odoo
- +1
Private client codebase
06
DEVtrust 2025 — Present
File-Parser
High-throughput Go tool that maps messy CSV and PDF data onto relational schemas without duplicating rows.
- Go
- PostgreSQL
- Supabase
Private client codebase
Experience
Where the six years went
From a first web development role automating sales reports to leading a web team and now building production AI systems. Expand any role to see what it actually involved.
What I did
- Engineered E-Abhilekh, a departmental document management platform on Django and GCP, with an AI smart-search layer over uploaded PDFs.
- Architected the Updesh RAG chatbot using FastAPI, Nuxt.js and Chroma DB, including a Docling-based parse, chunk and embed pipeline.
- Built the ATR order management service on AWS Chalice, Lambda, DynamoDB and S3, integrating third-party systems over GraphQL webhooks.
- Developed a real-time student management and chat platform on Django 5 with Redis, Celery and WebSockets.
- Wrote a high-performance Go tool mapping CSV and PDF data onto relational schemas with duplicate-safe imports.
Highlight
Took generative AI from proof of concept into departmental production — semantic search over document archives plus automated drafting of notices, circulars and summaries from file context.
Stack
- Python
- Django
- FastAPI
- Go
- Nuxt.js
- Qdrant
- Chroma DB
- Ollama
- GCP
- AWS Lambda
- DynamoDB
- Redis
- Celery
Expertise
What I actually mean when I say AI engineering
Every one of these is a thing I have had to make work under real constraints — private data that cannot leave the estate, documents that resist parsing, and users who will notice immediately when a model makes something up.
AI Engineering
systems that reason over your dataRetrieval-Augmented Generation
parse → chunk → embed → retrieve
The whole pipeline, not just the vector search. Documents are extracted with Docling, chunked so context survives the split, embedded into Qdrant or Chroma, and retrieved with the filters that keep one tenant out of another's data. Retrieval quality is a data problem long before it is a model problem.
- Qdrant
- Chroma DB
- Docling
- LangChain
Generative AI
drafting from context
Using models to produce real work product — notices, circulars and summaries generated from the contents of a specific file, in the format the department actually files.
- LLMs
- Ollama
- Prompt design
AI Agents
tools, not just text
Giving a model typed tools and letting it choose: query the database, call the API, fetch the document. The engineering is in the tool boundary and the failure handling.
- LangChain
- Function calling
Agentic AI
multi-step, stateful
Workflows that plan, act, observe and retry across several steps while holding state — graph-based orchestration rather than one prompt hoping to do everything at once.
- LangGraph
- State machines
LLM Integration
local or hosted
Serving models locally through Ollama when data cannot leave the estate, or calling hosted APIs when it can — behind one interface so the choice stays reversible.
- Ollama
- Streaming
- Token budgets
AI Guardrails
refuse over invent
A chatbot that answers "I do not have that record" is more valuable than one that guesses. Strict grounding to retrieved context, input and output validation, and scoping so answers cannot cross tenant boundaries.
- Grounding
- Schema validation
Evaluation with LangSmith
trace, score, regress
Tracing chains to see which retrieval step actually degraded, scoring answers against a fixed question set, and catching regressions before a prompt change ships.
- LangSmith
- Eval sets
Backend Architecture
the part that has to stay upAPI Architecture
contracts first
Typed, versioned, documented interfaces — FastAPI with Pydantic models, GraphQL where the client shape justifies it, and idempotent webhook handlers because third parties retry.
- FastAPI
- Pydantic
- GraphQL
- OpenAPI
Microservices & SOA
split on seams
Service boundaries drawn where the domain actually divides, with message brokers between them so one slow consumer degrades a feature instead of the product.
- RabbitMQ
- Celery
- Docker
Scalable Backend Systems
async by default
Long work moves to Celery, real-time traffic moves to WebSockets over Redis, and the database is asked the right question — window and aggregation functions instead of a loop in Python.
- PostgreSQL
- Redis
- Celery
- WebSockets
References
What it's like to work with me
Illustrative quotes drawn from the kind of feedback these projects generated — real named references available on request.
Dipu took the approval logic we had been patching for months and turned it into a configuration problem. New request types stopped needing a release.
He was the one insisting the chatbot should say it does not know. That decision is why the department trusted the tool enough to keep using it.
Reliable with a deadline and unusually careful about the database layer. As team lead he reviewed thoroughly without becoming a bottleneck.