AI Engineer · Agentic AI & LLM Systems · GraphRAG · Production MLOps

Shravan
M.

I build autonomous AI agents and retrieval systems and ship them to production — from data pipeline to a deployed, evaluated, observable, autoscaling service. 6+ years across pharma and commercial analytics, specializing in agent orchestration, tool calling, and GraphRAG over knowledge graphs and documents, deployed and operated on AWS.

View my work Get in touch
6+
Years experience
2M+
Graph nodes shipped
40%
TTFT reduction
8
Certifications
Core capabilities

What I build with

Production systems — not demos. Every skill listed has shipped to real users under real load.

AG
AI Agents & LLM Systems
Autonomous agents that decide when to search, call a tool, or answer. Agent loops and control flow, multi-agent orchestration, tool calling via MCP, agent memory, and grounded, traceable answers.
LangGraphLangChainMCP Tool callingMulti-agentRAG Structured output
GR
Graph & Retrieval
Knowledge graphs and GraphRAG pipelines. Hybrid retrieval combining semantic vector search with graph traversal, plus graph data science (GDS) for clustering and influence scoring fed back into ranking.
Neo4jGraphRAGCypher Neo4j GDSLouvain / PageRank pgvectorOpenSearch
EV
Evaluation & Quality
Golden-set evaluations, groundedness and hallucination checks, and answer-quality metrics. CI quality gates that block releases on regression, with full agent run tracing for debugging.
TruLensLangSmithGolden sets GroundednessCI quality gatesRun tracing
OP
Platform & MLOps
Deploy and operate AI systems on AWS. One build promoted dev → prod, horizontal scaling and autoscaling, semantic and KV caching, and cost-bounded serving with full observability.
AWS BedrockSageMakerECS / Lambda Kubernetes (EKS, Helm, HPA)Terraform GitHub ActionsPrometheus / Grafana
DA
Data & Big Data
Large-scale data platforms and ETL/ELT pipelines feeding analytics and ML. S3-to-Databricks ingestion, transformation at scale, and orchestrated nightly jobs.
DatabricksSnowflakeSpark / PySpark AirflowKafkaRedisPostgreSQL
CD
Languages & Backend
Python is home — async services, data, and ML. FastAPI with streaming for low-latency serving. SQL and Cypher for data and graphs.
PythonFastAPIAsync SQLCypherREST APIs
Career

Where I've shipped

Enterprise pharma environments — regulated, high-stakes, production-only. No sandboxes.

Jan 2025 — Present  ·  Hyderabad, India
Senior Analyst — Data Solutions
Sanofi
  • Built a LangGraph-based autonomous agent that orchestrates tool execution and GraphRAG workflows across multiple data sources and a knowledge graph — producing grounded, cited answers for 35 analysts across 5 teams and cutting research for critical commercial insights from ~2 business days to ~3 hours.
  • Served it through a FastAPI streaming service — first response in under ~1.5s, full answers in ~4s at p95.
  • Cut latency under load: ~40% on time-to-first-token and ~27% on p95, via streaming, prompt caching, and horizontal scaling on ECS.
  • Deployed on AWS — Bedrock for the model, containers on ECS, defined in Terraform and shipped through GitHub Actions; setup dropped from days to ~30 minutes.
  • Drove AI engineering standards: established the core enterprise GenAI framework now adopted by cross-functional teams, and ran internal workshops and alignment sessions.
  • Optimized unit economics — capped agent steps, read-only tools, semantic caching, and model routing — to ~$0.04 per question and ~45% lower model spend.
  • Built ML models for commercial analytics: customer and account segmentation (K-Means + graph clustering), feature engineering, and statistical analysis, with experiments tracked in MLflow.
  • Built the eval harness — a golden set scored for grounded correctness and honest refusal using TruLens with LangSmith tracing, run as a CI quality gate that blocks any release below ~90%.
  • Built the data foundation: S3-to-Databricks ETL pipelines transforming raw source data into clean, query-ready tables on Databricks and Snowflake.
LangGraphGraphRAGNeo4j MCPTruLensLangSmith FastAPIAWS BedrockECS TerraformDatabricksSnowflakeMLflow
May 2022 — Jan 2025  ·  Hyderabad, India
Data Engineer — Commercial Analytics
Novartis (via Sebone Technologies)
  • Engineered a GraphRAG agent navigating a Neo4j medical-expert graph (~2M nodes, ~8M relationships) to automate complex network analysis.
  • Ran graph data science on Neo4j — the GDS library for ML clustering and graph algorithms (community detection, Louvain, centrality, PageRank-style influence scoring) to surface the most influential experts and hidden clusters, and persisted outputs back to the graph as node properties to dynamically weight downstream hybrid-retrieval rankings.
  • Implemented hybrid retrieval — semantic vector search combined with graph traversal — lifting accuracy on relationship questions from ~60% to over 90% versus plain search.
  • Built the data pipeline on Databricks: S3-to-Databricks ETL ingesting, extracting, chunking, and embedding ~500K documents into the graph on a nightly Airflow schedule.
  • Hardened for production: deduplicated records, input sanitization, strict step-by-step traceability, and graceful degradation fallbacks.
Neo4jNeo4j GDSCypher GraphRAGDatabricksPySpark SnowflakeAirflowAWS
Oct 2020 — May 2022  ·  Bangalore, India
Data Engineer
Atrimed Pharmaceuticals
  • Built ETL pipelines in Python and Spark on Databricks — ingesting raw data into S3, cleaning and transforming it at scale, and landing it in a SQL warehouse that analysts and models could actually use.
  • Modeled the warehouse — turned messy source data into well-structured tables and the metrics the business asked for, with validation on the way in so downstream reports could be trusted.
  • Supported early ML work: prepared and engineered features from the warehouse for analytics and modeling, partnering with stakeholders to frame questions the data could answer.
  • Automated and scheduled production pipelines on Airflow, eliminating manual intervention and ensuring data reliability.
PythonSparkDatabricks S3SQL warehouseAirflow Feature engineering
Built

Projects & products

From pharma knowledge graphs serving field teams to founder-built SaaS — each one a real system under real production load.

01 / Enterprise · Pharma · Novartis
US Pharma KOL Pipeline
A GraphRAG agent over a Neo4j medical-expert graph — modeling US medical experts and their relationships so analysts could explore connections that flat search couldn't surface, and automate complex network analysis at scale.
  • Neo4j knowledge graph of ~2M nodes and ~8M relationships built from CRM, MDM, and third-party data sources
  • Neo4j GDS for ML clustering and graph algorithms (community detection, Louvain, centrality, PageRank-style influence scoring) — outputs persisted back to the graph as node properties to weight retrieval
  • Hybrid retrieval combining semantic vector search with graph traversal — accuracy on relationship questions lifted from ~60% to over 90%
  • S3-to-Databricks ETL ingesting, chunking, and embedding ~500K documents nightly on Airflow
  • Hardened for production: deduplication, input sanitization, traceability, and graceful degradation
Neo4jNeo4j GDSGraphRAGCypher DatabricksPySparkAirflowAWS
02 / Enterprise · Pharma · Sanofi
Commercial Insights Agent
A LangGraph-based autonomous agent for Sanofi commercial and pharma analytics teams — it decides when to search, when to call a tool, and when it has enough to answer, producing grounded, cited insights over multiple data sources and a knowledge graph. Built to be cheap, fast, and trustworthy in a regulated environment.
  • Agent orchestrates tool execution and GraphRAG workflows — used by 35 analysts across 5 teams, cutting research from ~2 business days to ~3 hours
  • FastAPI streaming: first response under ~1.5s, ~4s at p95; latency cut ~40% (TTFT) and ~27% (p95)
  • Eval harness with TruLens + LangSmith — golden-set scoring of groundedness and honest refusal, gating CI below ~90%
  • Unit economics: ~$0.04 per question and ~45% lower model spend via step caps, semantic caching, and model routing
  • Deployed on AWS Bedrock + ECS, defined in Terraform, shipped through GitHub Actions — setup days → ~30 minutes
LangGraphGraphRAGNeo4jTruLens LangSmithFastAPIAWS BedrockTerraform
03 / Enterprise · Pharma · Sanofi
French Pharma Disclosure Suite
A fully automated, AI-powered system for French pharmaceutical disclosure reporting — replacing a manual process end to end. An ingestion service feeds data into an audit engine that evaluates it against complex regulatory business rules, flags data-quality issues for review, and an autonomous agent assembles the final compliant reporting.
  • Ingestion service to process incoming data from multiple source systems
  • Audit engine evaluating data against complex regulatory business rules, flagging data-quality issues and exceptions for review
  • Autonomous agent assembling the final compliant reporting, with rules and validations built to extend as requirements change
  • Engineered on LangGraph, Snowflake, Databricks, and Neo4j behind a REST API, deployed on AWS
LangGraphSnowflakeDatabricksNeo4j REST APIAWS
04 / Founder
OpsGraph
A founder-built GraphRAG intelligence platform that answers business questions over a knowledge graph. A LangGraph agent pipeline turns plain-English questions into Cypher, runs them, synthesizes insight, and suggests an action — schema-driven and multi-tenant by design.
  • LangGraph agent pipeline: classify intent → generate Cypher → run query → synthesize insight → suggest action
  • Over Neo4j and Snowflake, served through FastAPI, deployed on AWS with Docker and Terraform
  • Multi-tenant by design — a schema-driven setup switches to a new client with a single config change, no code edits
  • Proactive engine surfacing nightly AI-generated operational alerts; Next.js dashboard with WebSocket streaming
FastAPILangGraphNeo4jSnowflake Next.jsPostgreSQLRedisDocker
05 / Founder
LinkGraph
Graph-native SEO intelligence platform. Ingests Screaming Frog crawl exports, computes PageRank on the internal link network, and lets SEO teams query their entire site structure in plain English — replacing hours of manual spreadsheet analysis.
  • Screaming Frog CSV upload → MERGE into Neo4j → PageRank computed automatically in-browser
  • D3 force-directed internal link graph — node size = PageRank, colour = page type
  • Cannibalization detection, orphan pages, quick-win keyword engine (positions 4–20)
  • Google Search Console OAuth integration for live keyword click and impression data
FastAPINeo4jD3.jsGPT-4o Next.js 14GSC APIDocker
06 / Open source
GraphKG
Dockerized Neo4j knowledge graph production starter — covering every pattern that bites engineers in their first production graph deployment. Ships with a companion 27-page technical book covering architecture, optimisation, and a 40-point pre-production checklist.
  • Procedural Python loader with idempotent MERGE patterns — safe to re-run on any dataset
  • Pure-Cypher PageRank + community detection — no GDS plugin required
  • 27-page PDF: graph modelling, 10 use-cases with full Cypher, performance tuning, production checklist
  • Batch loading strategy, constraint design, and index management documented end-to-end
Neo4jPythonDockerCypherReportLab
Credentials

Certifications

Every cert represents a real system I've shipped with that technology.

Neo4j Certified Professional
Neo4j & Generative AI (Graph Database and GraphRAG)
Microsoft Azure Fundamentals (AZ-900)
Snowflake Data Cloud Deployment Framework
Confluent Apache Kafka Fundamentals
Apache Airflow Fundamentals (Astronomer)
Databricks Lakehouse Fundamentals
Dataiku Core Designer

Let's build something
that actually ships.

Open to AI engineering roles building agentic AI and retrieval systems in production, freelance GraphRAG engagements, and founder conversations.

Shravan.thokottu@gmail.com linkedin.com/in/shravan-m12 opsgraph.in +91 81237 43491

Hyderabad, India · Available for remote worldwide