Professional Record

Production Delivery & Architecture

Venkata Naveen Busiraju · Gen AI Consultant at Atos

I build the unphotogenic half of enterprise AI.

Long-context retrieval over messy enterprise files that were never meant to be parsed. Fault-tolerant multi-agent state machines with strict schema enforcement. And the circuit breakers, priority queues, and token meters that keep generative systems standing when production traffic arrives.

Inspect Shipped Systems Direct Transmission
Work History & Enterprise Delivery
Dec 2023 – Present Current Role

Gen AI Consultant

Atos · London, UK
  • Built a long-context RAG system over course material — PDF, DOCX, PPTX, XLSX — that checks it against a data-quality checklist and files the compliance report to SharePoint. Review time down 65%.
  • Shipped an accounts-receivable email agent that classifies mail, pulls invoice data out of the attachments and validates it before anyone sees it. Resolution time down 55%.
  • Built the pipeline underneath: OCR with pdfplumber and pytesseract, Llama-3 prompting, then an agent validation pass. 96% field-level accuracy, ERP-ready JSON out.
  • Orchestrated it on LangGraph and MCP with separate retrieval, validation, synthesis and orchestration agents, schema enforcement and fault tolerance throughout.
  • Ran the delivery side too: Docker, Kubernetes and GitHub Actions onto GCP and AWS, plus Angular and Node.js migrations off the legacy architecture.
Nov 2019 – Jan 2022

Associate Consultant

Atos · Chennai, IN / Remote
  • Ran continuous integration on GCP with GitHub Actions, Docker and Kubernetes.
  • Moved legacy projects onto Angular and Node.js, and Dockerised microservices so development and production stopped disagreeing.
  • Wrote Java Spring Boot services on GCP for the decarbonisation project.
Naveen Busiraju

Venkata Naveen Busiraju

Gen AI Consultant at Atos

London, UK · Chennai, IN

Gen AI Consultant at Atos, five years in. I moved from cloud and full-stack delivery into building generative AI for enterprises, which in practice means the unphotogenic half: long-context retrieval over real documents that were never meant to be parsed, multi-agent workflows with schema enforcement and fault tolerance, and OCR-plus-LLM pipelines accurate enough to write into an ERP. I still build the stack around it — Java, Spring Boot, Node.js, Angular, Python — and ship it on GCP and AWS with Docker, Kubernetes and GitHub Actions.
Technical Matrix
Languages
PythonJavaNode.jsJavaScriptAngularSQL
Gen AI
LangChainLangGraphLlamaIndexTransformersHuggingFaceRAGASLiteLLMAzure OpenAI
ML
PyTorchTensorFlowScikit-LearnspaCyNLTKPandasNumPy
Platform
DockerKubernetesTerraformGitHub ActionsJenkinsRedisChromaDB
Cloud
Google CloudAWSAzure AI Search
Verified Cloud Credentials
AWS Certified DevOps Engineer – Professional
Amazon Web Services · Feb 2025 (Valid through Feb 2028)
Google Cloud Certified Associate Cloud Engineer
Google Cloud · 2021
Credential ID: ZVW1QW
Academic Degrees
MSc Computer Science
University of East London (Feb 2022 – Feb 2024)
BTech Computer Science and Engineering
Gudlavalleru Engineering College (Jul 2015 – Apr 2019)
Enterprise Validation

Systems Performance Under Load

Audited outcomes delivered across enterprise production workflows: reducing human verification bottlenecks, accelerating invoice pipelines, and enforcing OCR extraction fidelity.

65%
less review time

long-context RAG over course material, checked against a data-quality checklist

55%
faster AR resolution

email agent that classifies, extracts and validates invoice data

96%
field-level accuracy

OCR into Llama-3 into agent validation, ERP-ready JSON out

Systems Architecture Ledger

Eight Production Implementations

Every system below was designed to solve a specific production failure mode: model downtime, queue runaway, unverified tool execution, budget leakage, or unmeasured retrieval accuracy.

Idea in, running app out
Operational (Live)

Agentic App Builder

Describe an app in a sentence and four LLM agents — planner, developer, tester, reviewer — pass it around a LangGraph state machine until the code passes its own tests. The result boots in a sandboxed subprocess behind a reverse proxy that feeds browser console errors back into the loop. Five runtimes: Python, Node, React, Angular, static.

User Prompt → LangGraph: [Planner ⇄ Developer ⇄ Tester ⇄ Reviewer] → Sandboxed Subprocess (Port 8009)
LangGraphLangChainFastAPISSEDockerPython 3.12
Live System Repository
When the model you wanted is down
Operational (Live)

Model Fallback Router

Sits in front of your LLM providers. When one starts failing or breaches its latency SLA, a circuit breaker trips and traffic moves to the next model in the chain. The dashboard shows which model actually served each request.

Inbound Request → Circuit Breaker (Latency SLA Guard) → Primary ⚡ Fallback Model (200 OK)
LiteLLMFastAPICircuit BreakerDockerPython
Live System Repository
Backpressure before the meltdown
Operational (Live)

Async Inference Queue

A Redis-backed priority queue for LLM jobs. Workers pull by priority, backpressure kicks in before the queue runs away, and anything that exhausts its retries lands in a dead-letter queue you can actually inspect.

Producers → Redis Priority Queue [P1 | P2 | Backpressure] → Workers / DLQ Inspection
RedisFastAPILiteLLMDockerPython
Live System Repository
Every call priced and attributed
Operational (Live)

Token Cost Budget Dashboard

Prices each LLM call and attributes it to a team and a project. When a budget is spent the next request is refused rather than quietly billed. Works with any provider LiteLLM speaks to. 40 tests.

Model Token Stream → Pricing & Attribution Matrix → Budget Guard (Strict Limit Refusal)
LiteLLMFastAPISQLiteBudget EnforcementPython
Live System Repository
Did that retrieval change help?
Benchmark Suite

RAG Eval Suite

Scores a RAG pipeline on faithfulness, answer relevancy, context precision and recall with RAGAS, so a retrieval tweak stops being a matter of opinion. Local sentence-transformers embeddings, ChromaDB behind it, FastAPI in front.

Query + Document Context → ChromaDB Retrieval → RAGAS: Faithfulness, Relevancy, Precision, Recall
RAGRAGASChromaDBLiteLLMFastAPI
Repository
A ReAct loop that survives a flaky tool
Operational (Live)

Agent Tool Calls & Retries

A ReAct loop built on native function calling, with exponential backoff when a tool fails and a JSON trace of every step the agent took. Swap the model without touching the agent. Ships LangGraph, CrewAI and Google A2A adapters. 31 tests.

Prompt → ReAct Think ⇄ Tool Call (Exponential Backoff) → Structured JSON Step Trace
ReAct AgentLiteLLMLangGraphCrewAIFastAPI
Live System Repository
Lifecycle carbon for a laptop fleet
Operational (Live)

Carbon Calculator Platform

Works out the lifecycle carbon of a laptop fleet: ingests vendor spec data, applies country-level grid emission factors, and gives people comparison dashboards, with admin controls behind role-based access.

Vendor Specs → Country Grid Emission Factor Model → Lifecycle Comparison Dashboard
Node.jsExpressMySQLDockerGitHub Actions
Live System Repository
The roadmap, on a phone
Operational (Live)

GenAI Roadmap

A mobile-first study companion for working through the GenAI roadmap — structure and progress tracking on the device you actually have on you.

Curriculum → Foundations → Embeddings → RAG → Agents → Evals → Mobile Progress Tracking
GenAIMobile-firstEducation
Live System
Direct Transmission

Initiate an Engineering Inquiry

Open to Gen AI consulting — London / Chennai. If your enterprise requires production RAG pipelines, resilient multi-agent state machines, or resilient LLM proxy infrastructure, send a direct transmission.

[email protected] LinkedIn GitHub