Skip to content
Mehul Mangave

Mehul Mangave Cloud & AI Infrastructure Engineer

I build cloud infrastructure teams can ship on without provisioning it by hand.

I'm an infrastructure engineer based in California, working across AWS, Terraform, and Kubernetes, with recent work on RAG and knowledge-graph systems. I care about infrastructure that is version-controlled, cost-aware, and predictable to operate.

View projectsDownload résumé
San Jose, CaliforniaOpen to relocationGitHubLinkedInEmail

Impact in numbers

  • $500K

    Cut from annual AWS spend through rightsizing and scheduling

    Syngenta

  • 100+

    AWS accounts migrated onto version-controlled Terraform Cloud

    Syngenta

  • 99.9%

    Uptime sustained for more than 10,000 users

    Syngenta

  • 60%

    Manual deployment effort removed by Terraform automation

    State Street

About

Background, and what I'm looking for next.

I've spent two years on cloud infrastructure at Syngenta and State Street, mostly on the parts nobody demos: provisioning paths, module libraries, and the policy that runs in front of them.

Most of that is infrastructure as code. I've migrated account estates onto Terraform Cloud, written modules other teams build on, and put policy checks ahead of deployment so non-compliant resources never reach an account. The rest is cost and reliability, which usually means rightsizing instances and switching off environments nobody is using.

More recently I've been building on top of that plumbing rather than only underneath it, on RAG pipelines and knowledge-graph reasoning systems. I'd rather own infrastructure end to end than hand a Terraform plan across a wall.

Education

  • San Jose State University

    M.S. Computer Engineering

    May 2026 San Jose, CA

  • Vishwakarma Institute of Technology

    B.Tech. Electronics and Telecommunications Engineering

    May 2023 Pune, India

Experience

Two years running cloud infrastructure in production, from account migrations and cost work to policy-as-code.

  1. Jun 2025 – Aug 2025

    State Street Bank and Trust

    Cloud & AI Infrastructure Engineer Intern

    Boston, MA

    • Built a production RAG system over internal Confluence documentation with Python backend APIs, giving engineers self-serve answers and cutting repeat support questions to the platform team.
    • Engineered a Terraform automation proof of concept that generates infrastructure templates from user-defined requirements using in-house modules, removing over 60% of the manual effort in deployment workflows.
    • Worked with the Service Enablement team to architect, build, and document reusable Terraform modules for provisioning AWS services in line with organizational standards.

    FocusTerraform · AWS · Python · RAG pipelines · Vector embeddings · IaC automation

  2. Jun 2023 – Jul 2024

    Syngenta

    Cloud Engineer

    Pune, India

    • Migrated 100+ AWS accounts to Terraform Cloud, replacing manual provisioning with version-controlled IaC that eliminated configuration drift and gave every team a consistent, auditable, Git-reviewed deployment path.
    • Cut $500K in annual AWS spend by rightsizing over-provisioned EC2 and RDS instances, scheduling non-production environments to power down outside business hours, and decommissioning idle volumes, snapshots, and load balancers.
    • Built reusable Terraform modules and CI/CD deployment pipelines that cut deployment time by 40% and sustained 99.9% uptime for more than 10,000 users, letting application teams ship without provisioning infrastructure by hand.
    • Implemented Sentinel policy-as-code across AWS and GCP in Terraform Cloud, blocking non-compliant resources at plan time rather than in review, and reducing manual review effort by 70%.
    • Migrated DynamoDB tables across AWS accounts using on-demand backup and restore, validating record counts and throughput configuration after cutover to complete the transition with zero data loss and minimal downtime.

    FocusTerraform Cloud · AWS · GCP · Sentinel · CI/CD · Cost optimization · DynamoDB

  3. Aug 2022 – Jun 2023

    Syngenta

    Software Engineering Intern

    Pune, India

    • Provisioned cloud infrastructure for 20+ applications across EC2, S3, RDS, DynamoDB, API Gateway, Lambda, ECS, ECR, and ELB.
    • Planned and implemented Terraform modules that standardized infrastructure management across departments, now used by more than 50 teams as a framework for secure, repeatable cloud provisioning.
    • Built and documented a solution for AWS database services including DynamoDB, Redshift, DocumentDB, and Timestream, since adopted by other teams, improving performance by 30% and data processing efficiency by 25%.

    FocusAWS · Terraform · Lambda · ECS / ECR · DynamoDB · Redshift · API Gateway

Projects

Each one opens into what it solves, how it is built, and what came out of it.

Timeline
Hackathon · 2026
Role
Agent architecture, capture pipeline, dashboard
Stack
Guild.ai Agents · Python · ClickHouse · Next.js · OpenUI

View Magpie on GitHub

Reading for research generates dozens of copied snippets across tabs and PDFs, and almost none of them survive into notes.

  • A clipboard watcher classifies each captured snippet into a topic notebook and streams it to ClickHouse.
  • Four composed, single-purpose agents (classify, summarize, re-file, chat), each kept pure as text-in, text-out.
  • Database access is isolated to the watcher and API routes, so no agent touches storage directly.
  • A live dashboard surfaces per-notebook summaries alongside a chat assistant over everything captured.

OutcomeWon 1st Prize for Most Innovative Use of Agents.

Challenge
Agents return malformed output often enough that anything built directly on their responses will break mid-demo.
Approach
Server-side validation with a deterministic fallback for bad model output, a non-blocking capture queue so the watcher never stalls behind inference, and narrowly scoped table mutations.
Result
A bad agent response degrades to the fallback instead of taking the dashboard down with it.

Timeline
Aug 2025 – Dec 2025 · Team of 2
Role
Adaptive scoring and evaluation
Stack
Python · PyTorch · RAG · LLMs

View Dynamic Weight Learning for Self-RAG on GitHub

Standard Self-RAG pipelines weight every retrieved document the same, so a barely relevant passage pulls on the answer as hard as the one that actually contains it.

  • Retrieved passages are re-ranked by relevance to the query rather than accepted in retrieval order.
  • Each passage receives an adaptive influence weight in place of the uniform one used by the baseline.
  • Implemented in Python and PyTorch on top of an existing Self-RAG pipeline.
  • Evaluated across more than 1,000 queries against the static-weight baseline.

Outcome~18% gain in response accuracy and relevance across 1,000+ evaluations.

Challenge
Re-scoring every retrieved passage per query adds work to the hot path, so an accuracy gain is only worth having if it survives the latency it costs.
Approach
The weighting was folded into the retrieval scoring logic and optimized there, rather than added as a second pass over the candidate set.
Result
Accuracy and relevance improved by roughly 18% over the static-weight baseline with inference overhead kept low.

Timeline
Aug 2025 – Mar 2026 · Team of 4
Role
Graph traversal and reasoning-path generation
Stack
Knowledge Graphs · LLMs · Graph Traversal · Python

View Omnireasoner on GitHub

Multi-hop questions need several facts chained together, and a model answering from its own priors will produce a fluent chain that nothing actually supports.

  • Combines graph traversal over a 10K+ entity knowledge graph with LLM prompting.
  • Entity linking maps question terms onto graph nodes before any traversal begins.
  • Traversal emits an explicit reasoning path, so every answer carries the relations it came from.
  • Answers are grounded in retrieved graph relations rather than model priors.

Outcome~15% fewer hallucinated responses, with interpretable reasoning paths.

Challenge
A model asked to reason across several hops will quietly invent the missing link rather than report that it is missing.
Approach
Retrieved graph relations are supplied as the reasoning substrate and the traversal path is kept as part of the output, which makes the derivation inspectable instead of implicit.
Result
Hallucinated responses fell by roughly 15% in evaluation, and each answer ships with the path that produced it.

Timeline
Jan 2026 – Mar 2026
Role
Data model, REST APIs, component system
Stack
TypeScript · Next.js · Node.js · PostgreSQL · React

View QueueFlow on GitHub

Walk-ins and booked appointments compete for the same staff hours, and double-booking is the default failure mode.

  • Relational models for appointments, employee availability, customer records, and waitlists.
  • REST APIs that validate input and detect scheduling conflicts on write, so every client path is checked the same way.
  • Role-based workflows give customers, employees, and administrators separate views of one schedule.
  • A reusable responsive component set backs the booking flows and an admin analytics dashboard.

OutcomeRole-based scheduling with server-side conflict detection.

Challenge
Three roles read and write the same schedule, and any of them can create a conflicting booking.
Approach
Conflict detection sits in the API layer against the availability model rather than in the form, which keeps the rule in one place as new clients are added.
Result
Administrators track bookings, wait times, and staff utilization from a single dashboard on desktop and mobile.

Timeline
Aug 2025 – Dec 2025
Role
Multi-step flow, API integration, persistence, export
Stack
React · JavaScript · Material UI · Framer Motion · OpenAI API

View SnapTrip on GitHub

Planning a trip means reconciling destination, dates, interests, pace, and accessibility needs across a dozen open tabs.

  • A multi-step form collects destination, schedule, interests, pace, and accessibility needs before anything is generated.
  • Structured preferences post to a REST API that calls the OpenAI API and parses JSON responses back into itineraries.
  • Trips persist client-side in localStorage, and finished itineraries export to PDF with jsPDF.
  • Light and dark themes, responsive from small phones through desktop.

OutcomeLighthouse performance 97, accessibility 87.

Challenge
Model responses are slow and loosely structured, but the planner still has to feel immediate and let people revise a plan without starting over.
Approach
Preferences are held as structured state and sent as JSON, so regenerating a day is a targeted call rather than a fresh questionnaire.
Result
Travellers can regenerate part of a plan without re-entering anything, and take the result offline as a PDF.

Timeline
Jan 2022 – Jun 2022
Role
Feature extraction, classifier selection, real-time feedback
Stack
Python · SIFT · PCA · Machine Learning

A driver usually learns the car ahead is turning once it has already started, which is late when traffic is moving.

  • SIFT extracts features from images of the vehicle ahead.
  • Principal component analysis reduces those features before they reach a classifier.
  • Six classifier families were trained and compared on the same reduced feature set.
  • Predictions are delivered as real-time audio, so the driver never has to look away from the road.

Outcome95% prediction accuracy; 30% reduction in driver reaction time.

Challenge
SIFT produces a high-dimensional descriptor set, which is more than a classifier needs and more than a real-time loop can afford to carry.
Approach
PCA cuts the dimensionality first, and six classifier types were compared on the reduced set rather than committing to one up front.
Result
95% accuracy predicting turning direction, with audio feedback cutting driver reaction time by 30%.

Timeline
Aug 2021 – Dec 2021
Role
Signal processing and classification
Stack
ECG Signal Processing · SVM · scikit-learn · Python

Arrhythmia is often caught only after it has done damage, and the ECG that would show it earlier has to be interpreted before it is useful.

  • ECG signals are processed before classification rather than fed in raw.
  • A support vector machine classifies the processed signal into arrhythmia types.
  • Built with scikit-learn using classical techniques rather than deep models.

OutcomePresented at the 4th International Conference on Robotics, Intelligent Automation and Control Technologies, VIT Chennai.

Challenge
Raw ECG carries as much recording artifact as rhythm, and a classifier will learn whichever of the two is easier.
Approach
Signal processing runs ahead of classification so the model sees features of the rhythm itself, and classical methods were chosen over deep models to keep the result explainable.
Result
The work was accepted and presented at an international conference.

Open source

How I work

Three things that show up in most of what I build.

Infrastructure as code, with guardrails

At Syngenta I moved more than 100 AWS accounts onto Terraform Cloud, which replaced manual provisioning with a Git-reviewed path every team could audit. Sentinel policy-as-code then ran at plan time across AWS and GCP, so a non-compliant resource was blocked before it existed rather than caught in review afterwards, cutting manual review effort by 70%.

Cost and reliability are one problem

Taking $500K out of annual AWS spend was unglamorous work: rightsizing over-provisioned EC2 and RDS instances, powering down non-production environments overnight, and deleting volumes, snapshots, and load balancers nobody had claimed. The same visibility that finds waste finds fragility, and the platform held 99.9% uptime for more than 10,000 users throughout.

Applied AI needs solid plumbing

The RAG system I built at State Street answers questions from internal Confluence documentation through Python backend APIs, and cut repeat support questions to the platform team. The model was never the hard part. Ingestion, retrieval quality, and the infrastructure underneath decided whether the answers could be trusted.

Toolkit

What I reach for day to day. Hover or focus a row to hold it still.

  • AWS EC2
  • S3
  • RDS
  • DynamoDB
  • Lambda
  • ECS / ECR
  • API Gateway
  • ELB
  • GCP
  • Azure
  • Kubernetes
  • Docker
  • Terraform
  • Terraform Cloud
  • Sentinel
  • Policy as code
  • CI/CD
  • CloudWatch
  • Prometheus
  • Grafana
  • RAG pipelines
  • Vector embeddings
  • Knowledge graphs
  • AWS Bedrock
  • Vertex AI
  • PyTorch
  • TensorFlow
  • Python
  • Go
  • Java
  • C/C++
  • Bash
  • SQL

Let's build something useful.

I'm looking for full-time cloud, infrastructure, and platform engineering roles. If you're hiring, or you just want to talk through a problem, my inbox is open.

mehul.mangave@sjsu.eduRésumé (PDF)