Skip to content

Open to AI/ML and full-stack roles

Divyansh Shukla, AI engineer · RAG & agentic systems

AI engineer · RAG & agentic systems

I architect production retrieval and agent systems on Azure, end to end.

Ingestion pipelines, hybrid and agentic retrieval, LLM evaluation, and the network-isolated cloud infrastructure underneath, plus the full-stack apps people actually use.

years shipping software
5+
systems deployed to production
8
projects delivered
18
generations of one production RAG platform
4
  • Azure OpenAI
  • Azure AI Search
  • Agentic retrieval
  • AutoGen
  • Python · FastAPI
  • TypeScript · React
  • Node.js
  • Azure Functions
  • Private networking

(01) Selected work

Systems that answer from evidence.

Retrieval, agents and the platforms around them, from first prototype to production.

FlagshipProduction since 20242024–2026

Public-Records AI Assistant

Lead engineer · retrieval, ingestion, evaluation & infra

Architected and evolved a citation-grounded RAG assistant for a public-sector authority across four generations, from hybrid search with LLM query rewriting to Azure AI Search agentic retrieval over two federated indexes, driven by offline evaluation of production traffic.

production turns evaluated
416
substantive answers carrying citations
93.7%
architecture generations since 2023
4
  • Evaluation-driven re-architecture. Built an offline evaluator over 416 production turns: regex failure taxonomy, citation parsing and a gpt-4.1 relevance judge. It exposed false rejections, follow-up and typo failures, and drove the move to agentic retrieval.
  • Federated agentic retrieval. Chose push-based ingestion into two schema-specific indexes federated by one knowledge base over a single indexer pipeline, so each source owns its schema and deletion strategy while queries still span one corpus.
  • Idempotent, fail-loud ingestion. Deterministic base64 parent IDs and indexed chunk IDs make re-ingestion delete-then-upsert with no orphans. Validation runs before deletes, OData filters are escaped, and vector-length checks fail loudly on a wrong embedding model.
Read the case study

Architecture

  1. 01

    Ingest

    Blob-triggered Python Function cracks disclosure PDFs with Document Intelligence, strips page noise, stitches cross-page tables and extracts reference numbers and letter dates.

  2. 02

    Crawl

    Containerised crawl4ai/Playwright deep crawler harvests the public website and linked documents, with incremental sync and guarded deletion reconciliation.

  3. 03

    Index

    Structure-aware token chunks embedded with text-embedding-3-large into two purpose-built HNSW indexes, upserted idempotently by deterministic parent and chunk IDs.

  4. 04

    Retrieve

    An Azure AI Search knowledge base plans subqueries over the last six turns, federates both indexes, then reranks and merges results with references.

  5. 05

    Generate

    A FastAPI backend grounds the model on the retrieved block and streams cited bullet answers over WebSockets, linking exact PDF pages or web URLs.

  6. 06

    Observe

    Turns logged to Cosmos DB by verified Entra role, then scored offline by regex failure classifiers and an LLM judge.

Evolution

  1. 2023 · Prototype

    Document Q&A demos on early Azure OpenAI

  2. 2024 · Production RAG

    Hybrid search, LLM query rewriting, queue ingestion, web crawler

  3. 2025 · Agentic

    Tool-calling agent, WebSocket streaming, private networking

  4. 2026 · Agentic retrieval

    KB query planner over federated indexes, LLM-judge evals

(02) How I build

From raw documents to grounded answers.

The engineering behind the demos: what it takes to run LLM systems for real users.

Documents in, structure preserved.

Retrieval quality is decided at ingestion. I build pipelines that keep the shape of a document, not just its text.

  • Layout-aware parsing with Document Intelligence, cached per document
  • Tables stitched across pages and split on row boundaries with repeated headers
  • Section-scoped, token-aware chunking with heading breadcrumbs
  • Idempotent parent/chunk upserts: no orphans, no duplicates on re-ingest
  • Incremental sync from Microsoft Graph delta feeds and deep web crawls
  • Queue-driven serverless workers with checkpointing and guarded deletes

(03) Experience

From sensors to agents.

5+ years, from IoT platforms to agentic AI. Roles and employers are on LinkedIn.

  1. 2026 — now

    Current focus

    Agentic retrieval & evaluation-driven engineering

    Re-architecting production RAG around Azure AI Search knowledge bases, with measurement before change.

    • Conversation-aware query planning over federated, purpose-built indexes
    • Structure-aware ingestion: layout parsing, table stitching, section-scoped chunking
    • Crash-safe Microsoft Graph delta sync into a compliance-grade search platform
    • LLM-as-judge evaluation of production traffic to drive retrieval redesigns
    • Agentic retrieval
    • Evals
    • Microsoft Graph
    • Container Apps
  2. 2025

    Multi-agent systems

    Tool-calling agent teams that route, draft, validate and review, streamed live to the browser.

    • AutoGen AgentChat teams behind FastAPI with token streaming over WebSockets
    • Production switch-over of a public RAG assistant to an agentic architecture
    • Merged public and admin portals behind private networking and an application gateway
    • AutoGen
    • FastAPI
    • WebSockets
    • Entra ID
  3. 2024

    RAG in production

    Owned retrieval, ingestion and releases for document Q&A systems running in dev and prod.

    • Hybrid keyword + vector retrieval with semantic reranking and cited answers
    • Queue-driven serverless ingestion with OCR, layout analysis and token-aware batching
    • Multi-tenant document Q&A API with OCR ingestion and vector search
    • Azure AI Search
    • Azure OpenAI
    • Azure Functions
    • Cosmos DB
  4. 2023

    Applied AI services

    First LLM applications and a portfolio of Azure AI services integrations.

    • Early Azure OpenAI chat and document Q&A prototypes
    • Vision, speech, language and form-recognition services behind a Node API
    • Azure OpenAI
    • Form Recognizer
    • LangChain
  5. 2021 — 2022

    IoT platforms & digital twins

    Full-stack platforms for monitoring buildings, warehouses and industrial assets in real time.

    • React/TypeScript dashboards with Node/Express APIs over SQL Server and Cosmos DB
    • Device simulators streaming into IoT Hub, with serverless digital-twin graph sync
    • React
    • Node.js
    • IoT Hub
    • Digital Twins

(04) Toolkit

The stack behind the answers.

01

AI & retrieval

  • Azure OpenAI (GPT-4o, 4.1, 5-series)
  • RAG architecture
  • Azure AI Search · hybrid + semantic
  • Agentic retrieval / knowledge bases
  • AutoGen multi-agent
  • Embeddings & vector indexes (HNSW)
  • Document Intelligence / OCR
  • LLM evaluation · LLM-as-judge
  • Prompt & tool design

02

Backend

  • Python · FastAPI · Quart
  • Node.js · Express
  • WebSockets & SSE streaming
  • REST API design
  • Microsoft Graph
  • Cosmos DB · SQL Server
  • Blob · Queue · Table storage
  • pytest · in-memory fakes

03

Frontend

  • TypeScript
  • React 19 · Vite
  • Angular (10 → 21)
  • Redux · MobX
  • Material UI · Fluent UI
  • Data viz · amCharts · Recharts
  • Azure Maps
  • MSAL / Entra ID sign-in

04

Cloud & infrastructure

  • Azure Functions · App Service
  • Container Apps
  • Private endpoints · VNet integration
  • Application Gateway
  • Managed identity · RBAC
  • Entra ID
  • Application Insights
  • IoT Hub · Event Grid · Digital Twins

(05) Ask my portfolio

Don't read it. Ask it.

An assistant grounded in my work. It answers with citations, checks my calendar and books a meeting, with your confirmation at every step.

Prefer your own assistant? Connect Claude, ChatGPT or Cursor to it.

Full-screen chat

MCP server

Connect your AI to this portfolio

Your own assistant can search my work, read case studies and see when I'm free. Read-only, no sign-in.

https://mcp.divyanshshukla.dev/mcp
  1. 1Open Customize → Connectors.
  2. 2Choose Add custom connector and paste the server URL.
  3. 3Pick No sign in, then save.