Applied AI engineer. Retrieval, agents and the platforms around them.
Azure AI Engineer and Team Lead with 15+ years building enterprise-scale platforms on Azure. Specialized in RAG architectures, agentic workflows and LLM-powered systems, delivering production-grade copilots and document intelligence pipelines. Today I build agents that do real work for engineers: they read documents, find the evidence, drive CAD tools and know when to ask a person. Most of my fifteen years of code was written by hand.
Role
Applied AI Engineer and Team Lead
Focus
Agentic AI and multi-agent systems, MCP tools, RAG copilots, document intelligence
Stack
Azure OpenAI, Azure AI Foundry, Azure AI Search, Python, C# and .NET, React and TypeScript
Leadership
Teams of up to 20 across the Bay Area and India; design reviews, coding standards, secure code
Based in
Santa Clara, CA (San Francisco Bay Area)
Education
M.S. Software Engineering, San José State University
The citadel beside this text is a map of the site. The keep in the middle is me; the camera moves to each building as you read.
Work, with my team, 2026. The newest work, in the engineering quarter between the keep and the gate.
An agent platform that works inside the CAD tools mechanical engineers already use
Plain English
Mechanical engineers spend their day in CAD and product-data tools, not in chat windows. AgentMesh brings AI to them: engineers open AISD, a portal my team built that runs inside Creo Parametric's own browser panel, and ask for help with the design in front of them.
AgentMesh sends each request to the right specialist agent, from CAD changes and enclosure experiments to battery comparisons, and that agent does the work in the engineer's own tools.
Technical
AISD's agents run in Azure. MCP servers run on the engineer's desktop beside the CAD session (a Creo Parametric MCP, plus Fusion MCP tools), so an agent in the cloud can read and drive the model locally.
Specialist agents and tools share the platform: the Creo CAD assistant, the Enclosure Agent, analysis tools, C-bucket tools and battery tools, alongside other agents in the wider system such as VASIM for virtual simulation.
Engineer in Creo
AISD portal
AgentMesh router
Specialist agent (Azure)
Desktop MCP
CAD / Windchill
Azure
MCP
Creo Parametric
Windchill REST API
Fusion
GitHub Copilot
Python
2.1Creo CAD assistant
AgentMesh, 2026
Ask for a CAD change in plain language, inside Creo
Plain English
Engineers describe what they want done to the model and the assistant carries it out in the open Creo session, without leaving the CAD tool.
Next, it will check each requested change and suggest better options before applying it.
Technical
The AISD agent runs in Azure and calls a Creo Parametric MCP server on the engineer's desktop, which performs the operations in the live session.
I'm extending an MCP that an earlier developer started, prioritizing missing capabilities, then adding validation and design suggestions to the CAD interaction.
Request in AISD
Agent plans steps
Creo MCP
Model updated
Validate + suggest
Creo Parametric
MCP
Azure
AISD
Runs
Inside Creo
2.2Enclosure Agent
AgentMesh, 2026
Explore an enclosure design before anyone builds it
Plain English
Engineers try different enclosure parameters and the agent runs the experiments for them, predicting how stiff each design will be.
It plots material weight against stiffness, so the best trade-offs are visible at a glance.
Technical
The agent runs parameter sweeps on a pre-trained Python surrogate model (deployed as-is), returning predicted stiffness for each variant.
Results come back as a weight-versus-stiffness Pareto chart.
Design parameters
Parameter sweep
Surrogate model
Stiffness
Pareto chart
Python
Surrogate model
AgentMesh
Output
Pareto front
2.3Battery tools
AgentMesh, 2026
Compare laptop battery designs across versions
Plain English
Battery decisions depend on how today's design compares with earlier ones. The battery tools pull those versions together so engineers can compare them while they design.
It is the platform's current focus, aimed at helping engineers make design decisions.
Technical
AISD queries the Windchill REST API for battery design data across versions, while Creo talks to Windchill directly for CAD operations.
Question in AISD
Windchill REST API
Battery versions
Comparison
Windchill
REST API
Creo Parametric
AISD
Data
Windchill
2.4C-bucket tools
AgentMesh, 2026
Laptop C-cover calculations, one prompt away
Plain English
The C-bucket is the top of the laptop base: the keyboard and palm-rest enclosure. Engineers can run its calculations straight from GitHub Copilot instead of hunting for a separate desktop tool.
Next, it becomes a web app like the other agents, with faster calculations.
Technical
A GitHub Copilot MCP integration launches a local UI tool and runs the C-bucket calculation on the machine's CPU or GPU.
Planned: move it from a desktop app to a web app and cut the time each calculation takes.
Copilot
MCP tool
Local UI
CPU/GPU calculation
GitHub Copilot
MCP
Desktop app
Runs
From Copilot
2.5Analysis tools
AgentMesh, 2026
Engineering analysis the agents can call
Plain English
Analysis tools that the platform's agents call, so their answers rest on calculations rather than guesses.
Technical
Exposed to AgentMesh agents as tools alongside the CAD and product-data integrations.
Agent
Analysis tool
Result
AgentMesh
MCP
2.6Fusion MCP tools
AgentMesh, 2026
The same agents, working in Fusion
Plain English
MCP tools that let AgentMesh agents work with Fusion as well as Creo, so the platform isn't tied to one CAD tool.
Technical
MCP servers that expose Fusion operations to the platform's agents, in the same pattern as the Creo Parametric MCP.
Agent
Fusion MCP
Fusion
Fusion
MCP
AgentMesh
03 / 09
New: agents I build
Personal systems from this year. I design them and build them with AI coding agents doing much of the typing, and they run every day on free infrastructure. They live in the workshop quarter, east of the keep. Read them in plain English or technical with keys 1 and 2.
3.1Deals Galore
Personal, open source, 2026
A shopping agent that tells me when a deal is real
Plain English
Every hour it checks more than fifteen deal sites, brand stores and forums for the things on my list, decides whether each deal is genuinely good, and sends me a Telegram message when one is worth buying.
It runs entirely on free services, and adding a new item to watch is one tap on its dashboard.
Technical
A scheduled GitHub Actions pipeline: RSS, Reddit, Shopify product JSON and schema.org pages feed a keyword prefilter, then batched LLM scoring on Groq with Gemini as fallback, with model discovery and per-model rate-limit handling.
Deterministic guards run after the model (budget range, brand price floors that catch fakes, per-store caps), so a confident model can't push a bad deal through. A Watch button opens a GitHub issue that a workflow merges into the config.
A research agent that reads the new AI papers every morning
Plain English
Hundreds of AI papers come out every day. Each morning Paper Desk reads the new ones and the big labs' blogs, keeps what matters for building RAG and agent systems, and writes a short summary of each.
It also notes why each paper might matter for my work, so I can decide in seconds what to read.
Technical
Pulls Hugging Face Daily Papers, the arXiv API and four lab RSS feeds, merges duplicates by arXiv id, and keeps candidates with per-topic keyword hits or high community upvotes.
An LLM scores relevance against a written reader profile and returns a TL;DR and a 'for you' line as JSON; keyword ranking takes over if the free quota runs out. The page is rebuilt into this site every day.
Skills that teach GitHub Copilot to build AI on Azure properly
Plain English
Coding assistants are quick to write AI features and quick to get them subtly wrong. This pack gives GitHub Copilot step-by-step playbooks for RAG, agents, evaluation and security on Azure, and Copilot loads the right one when the task needs it.
It also includes a scored catalog of open-source AI tools to help choose libraries.
Technical
23 Agent Skills (SKILL.md) installable as a Copilot CLI plugin or into .github/skills, organized along one lifecycle: choose, ingest, retrieve, orchestrate, prove, secure, operate.
Skills enforce keyless Entra ID auth, the Responses API, hybrid retrieval with security trimming, least-autonomy agent design, and evals before claims. Backed by a catalog of 82 OSS tools with verified metrics and licenses.
Production systems I designed and wrote with my teams, from 2012 to today: the architecture, the code and the late-night fixes were ours. They make up the old town, west of the keep.
2025 - Present
Agent Assist - Document Intelligence Platform
High-volume enterprise documents needed multilingual OCR, classification, extraction, validation, and summarization at scale without overwhelming human agents.
Solution
Built a multi-agent Azure AI Foundry workflow that processes 70K+ documents per month through routing, validation, and review checkpoints.
Architecture
Document ingress, OCR, language detection, classifier, extraction agents, human review queues, and analytics feedback loops.
Challenges
Latency control across multi-step pipelines
Cost-aware OCR and LLM usage
Reliable validation for imperfect documents
Lessons learned
Treat evaluation as product infrastructure
Use routing and confidence thresholds before invoking expensive models
Design agent workflows around clear validation contracts
Azure AI Foundry
Document Intelligence
Azure OpenAI
Python
Logic Apps
Functions
Dynamics CRM
2024 - Present
Crystal RAG + Copilot Platform
Knowledge workers needed grounded answers across Teams, CRM, and Office apps with enterprise controls and low-latency retrieval for 2K+ active users.
Solution
Delivered a secure Azure RAG platform with hybrid search, grounded citations, and feedback loops for CRM, Teams, and Office experiences.
Architecture
Content ingestion, enrichment, Azure AI Search, re-ranking, prompt orchestration, secure API tier, and Teams/Office frontends.
Challenges
Permission-aware retrieval
Prompt drift
Grounded citation quality
Feedback visibility across multiple data sources
Lessons learned
Retrieval quality beats prompt cleverness
Caching should be scoped by identity and freshness requirements
Analytics need to be part of the copilot product loop
Azure OpenAI
Azure AI Search
Azure AI Foundry
Python
Teams
Office
React
Applied Materials
EPG Taiwan Plant Enablement
A new Taiwan plant needed plant-specific workflows, data migration, and reporting support without disrupting existing integrations.
Solution
Led backend and data workstreams to extend .NET APIs, SQL Server schemas, and migration flows for Taiwan-specific process rules.
Architecture
.NET 6 API extensions, SQL Server schema changes, data validation workflows, migration support, and stakeholder sign-off loops.
Challenges
Plant-specific rule differences
Data integrity before go-live
Maintaining compatibility with existing integrations
Lessons learned
Process engineers should validate workflows early
Schema extensions need explicit consistency standards
.NET 6
SQL Server
Web API
Data Migration
Process Engineering
Applied Materials
Almanac Process Engineering Analytics
Process engineers needed a single view of operational data across multiple platforms to analyze trends, compare results, and create charts and reports.
Solution
Built and supported an analytics experience that integrated plant data into Angular, D3.js, and .NET Web API workflows for faster engineering review.
Architecture
Angular application, .NET 6 API, SQL Server and MongoDB-backed data views, D3.js visualization layer, and client validation workflows.
Challenges
Combining data from different systems
Interactive chart performance
Client-facing support and validation
Lessons learned
Analytics tools need fast feedback from domain users
Visualization work benefits from clean API boundaries
.NET 6
Angular
D3.js
SQL Server
MongoDB
Enterprise platform rebuild
Unified Action Tracker
Large teams were using fragmented ADO workflows that slowed visibility, ownership, and reporting across global operations.
Solution
Led a greenfield React and .NET multi-tenant platform that replaced scattered workflows with a dedicated user experience and analytics dashboards.
Architecture
React application, .NET API, optimized data model, migration tooling, Redis-backed low-latency paths, and dashboards.
Challenges
Migrating 10K+ records with zero data loss
Global reliability expectations
Multi-tenant data boundaries
Lessons learned
Migration observability is as important as the migration script
Workflow products need strong defaults and escape hatches
React
.NET
SQL
Redis
Azure App Service
CI/CD
Microsoft contract
Microsoft 365 Admin Center Feedback Platform
Product teams needed reliable, low-friction feedback capture inside a globally used admin experience for enterprise IT administrators.
Solution
Developed feedback elicitation components and surveys that surfaced user insights in the M365 Admin Center and informed product improvements.
Architecture
React and .NET feedback components, SQL-backed feedback flows, Azure deployment pipelines, and product analytics handoff.
Challenges
Global product reliability
Non-intrusive feedback UX
Turning submissions into actionable product signals
Lessons learned
Feedback capture should respect user context
R&D spikes are useful when product behavior is still forming
.NET 5
React
Redux
SQL Server
Redis
Azure
Teams
Microsoft contract
Tenant Touchpoints
Sales teams needed a dependable way to track tenants, Microsoft service usage, tasks, comments, and adoption views across a growing customer base.
Solution
Delivered React and .NET features for dashboards, task management, and Teams-connected workflows, helping adoption grow from 30 to 5K+ users in two years.
Architecture
React and Redux frontend, .NET service layer, SQL Server data model, Azure hosting, and Teams notification exploration.
Challenges
Daily sales workflow reliability
Stakeholder triage
Feature roadmap feasibility
Lessons learned
Internal tools need product-level support habits
Technical feasibility work can shape roadmap decisions
.NET 4
React
Redux
SQL Server
Azure
Azure DevOps
Teams
Cloud architecture case study
Multi-Region Order Processing Platform
A customer-facing order system needed resilient, real-time order status across regions.
Solution
Architected a multi-region AWS application with load-balanced ECS services and centralized managed RDS.
Architecture
Web app, API Gateway, ECS services, IAM controls, RDS MySQL, S3 assets, CloudWatch observability, and regional routing.
Challenges
Cross-region consistency
Access control design
Operational visibility
Lessons learned
Use managed primitives where they simplify recovery
Model order state transitions explicitly
AWS ECS
RDS MySQL
API Gateway
S3
Lambda
IAM
CloudWatch
CitiusTech
Provider Link Healthcare Platform
Home care and hospice providers needed dependable workflows for billing, revenue management, OASIS review, and clinical compliance.
Solution
Improved a healthcare full-stack application by fixing functional issues, optimizing performance, and raising code quality; exception rates were reduced by roughly 98%.
Architecture
.NET MVC application with AngularJS and jQuery client workflows, SQL Server persistence, revenue and compliance modules, and shared quality tooling.
Challenges
Legacy code maintainability
Performance bottlenecks
Clinical and billing workflow complexity
Lessons learned
Healthcare apps need careful regression control
Code quality work compounds when paired with production bug fixes
.NET MVC
AngularJS
JavaScript
jQuery
SQL Server
Code Quality
Rolta India
Department of Industries Digital Portal
Small-scale industry processes for Maharashtra Government needed to move from manual submissions to an online registration, claims, and certification workflow.
Solution
Designed and built front-end workflows for a public-sector portal that digitized submissions, document handling, and benefit processing.
Architecture
.NET MVC portal with jQuery and KnockoutJS views, SQL Server workflows, SharePoint integration, document upload paths, and approval-oriented screens.
Challenges
Digitizing government workflows
Document-heavy user journeys
Coordinating across a large delivery team
Lessons learned
Public-sector portals need simple task flows
Upload and certification steps need clear state models
.NET MVC
KnockoutJS
jQuery
SQL Server
SharePoint
Kotak Mahindra Mutual Fund
Sales Target and Achievements Portal
Sales teams needed a pan-India system to upload targets and achievements, review performance, and capture manager comments.
Solution
Designed and built a custom .NET MVC portal that supported target uploads, manager review, and reporting for a distributed sales organization.
Architecture
.NET MVC application with LINQ data access, jQuery interactions, SQL Server storage, target upload workflows, manager review, and reporting views.
Challenges
Requirement discovery with business users
Upload validation
Manager review and comment workflows
Lessons learned
Operational reporting apps live or die by data validation
Customer-facing requirement work improves delivery speed
.NET MVC
LINQ
jQuery
SQL Server
Reporting
Angel Broking
Angel Eye Trading Portal
A public single-page trading and portfolio management app needed functional fixes, change requests, and performance optimization.
Solution
Delivered fixes and enhancements across a .NET MVC and JavaScript application while improving responsiveness for customer-facing trading workflows.
Architecture
.NET MVC backend, LINQ-based data access, jQuery front-end interactions, SQL Server persistence, and SPA-style portfolio and trading screens.
Challenges
Customer-facing reliability
Performance-sensitive interactions
Frequent change requests
Lessons learned
Trading interfaces demand fast feedback loops
Small performance fixes matter in heavily used workflows
.NET MVC
LINQ
jQuery
SQL Server
SPA
Paramatrix Technologies
Enterprise Operations Portals
Multiple businesses needed custom internal systems for training, certification, client records, investment reporting, point-of-sale, inventory, and operational reports.
Solution
Built and maintained full-stack .NET MVC applications including Process Tracker, PIE, and Gold POS across finance, certification, CRM-style, and retail operations domains.
Architecture
.NET MVC portals with LINQ, jQuery, SQL Server schemas, reporting screens, process workflows, inventory modules, and deployment support.
Challenges
Serving different business domains
Balancing new features with production fixes
Keeping reporting and transactional data consistent
Lessons learned
Domain modeling transfers across line-of-business apps
Reliable deployments are part of full-stack ownership
.NET MVC
LINQ
jQuery
SQL Server
Inventory
Reporting
Legacy systems project
Distributed File Storage and Sharing
Explore distributed storage, replication, leader election, and work stealing in a clustered file service.
Solution
Built a Java and Netty-based distributed file sharing system with command queues, work queues, replication, and client messaging.
Architecture
Client connection layer, server state manager, command/work handlers, leader election, data replication, and queue orchestration.
Each employer is a tower on the wall, oldest beside the gate.
Dec 2018 - PresentTeam Lead / AI EngineerTech MahindraMicrosoft and Applied Materials contract programs
Led current Microsoft and Applied Materials contract programs spanning Azure AI, RAG, copilots, analytics, and enterprise platform modernization.
Designed a multi-agent document intelligence system processing 70K+ documents per month across OCR, multilingual classification, translation, extraction, summarization, and validation.
Led RAG and copilot delivery for CRM, Teams, Office, and React experiences serving 2K+ active users.
Engineered agentic workflows orchestrating multi-step LLM reasoning chains, cutting manual effort and turnaround time at scale.
Managed engineering teams of up to 20 people across the San Francisco Bay Area office and an offshore team in India, running design reviews, unblocking delivery, and mentoring engineers on React, .NET, Azure, and AI patterns.
Led initiatives across the company to reduce technical debt, improve application performance, and introduce coding standards, secure coding practices and a code review process.
Developed scalable React and .NET applications, Teams extensions and copilots, growing adoption from 30 to 5,000+ users.
Owned Azure infrastructure across App Services, Functions, Logic Apps, Key Vault, Redis, Blob Storage, Azure AI Search, Azure AI Foundry, and CI/CD pipelines.
Event-driven Azure platform with API orchestration, retrieval pipelines, secure document processing, and telemetry-backed feedback loops.
Designed chatbot integrations for REST services accessible from email, Microsoft Teams, and SMS.
Delivered workflow automation and platform integrations across systems including Workday and ServiceNow.
Conversational integration layer over enterprise workflow APIs.
Bots
REST APIs
Teams
ServiceNow
Workday
Aug 2015 - Jul 2016Senior Software EngineerCitiusTechHealthcare software platforms
Built and stabilized healthcare provider applications across billing, revenue management, OASIS review, and clinical compliance workflows.
Led architecture and exception-handling improvements that reduced exception rates by roughly 98%.
Improved performance and code quality in a legacy .NET MVC, AngularJS, jQuery, and SQL Server application stack.
Stabilized service and application layers with consistent error handling contracts.
.NET MVC
AngularJS
jQuery
SQL Server
Enterprise Web Apps
Nov 2014 - Jul 2015Software Engineer L2Rolta IndiaGovernment digitization
Built front-end and workflow functionality for a Maharashtra Government Department of Industries portal.
Helped digitize registration, document upload, certification, and approvals for small-scale industry processes.
Guided team members through implementation issues across .NET MVC, Knockout, jQuery, SQL Server, and SharePoint.
Transactional web systems for secure public-sector workflows.
.NET MVC
KnockoutJS
jQuery
SQL Server
SharePoint
Aug 2012 - Nov 2014Software EngineerParamatrix TechnologiesEnterprise operations
Delivered finance, trading, certification, reporting, CRM-style, and point-of-sale applications across multiple client domains.
Built full-stack .NET MVC portals using LINQ, jQuery, SQL Server, reporting workflows, and customer-facing requirement discovery.
Handled deployments, production fixes, performance optimization, and feature delivery alongside cross-functional teams.
Line-of-business applications backed by relational data models and reporting workflows.
.NET MVC
LINQ
jQuery
SQL Server
Reporting
06 / 09
Systems, with the tradeoffs shown
Pick a system, then a stage, to see why each part was chosen, what it costs and how it is secured.
Permission-aware retrieval with hybrid search, re-ranking, citations, and answer evaluation.
Stage 1 of 4
Identity Gateway
Why this
Enterprise search without ACL trimming leaks data.
Alternatives considered
App roles, Graph connectors
Pros
Central policy point
Cons
Requires careful cache scoping
Performance
Token validation and permission lookups must be cached carefully.
Cost
Identity lookups can dominate latency if not batched.
Security
Use Entra ID, Key Vault, managed identities, and per-user audit trails.
07 / 09
What colleagues say
Ashutosh led a vendor team of 4 people to work on a challenging web project, reporting directly to me. The project was to build a web SDK in React with both front-end and back-end components. His excellent work not only contributed tremendously to the success of the project, but also helped me offload some of my responsibilities. He would be a great asset in any organization.
Ashutosh is a great co worker to work with. Help other people in the team when they are struck. Takes full ownership of work that he is working. Always thinks future before writing code. Always follows best practices.
Ashutosh is an extremely enthusiastic and motivated developer, who is passionate about technology and is always up-to-date with the latest trends. He picks up new technology easily and is very good at training others how to use it. It was always a pleasure to work with him as he is friendly, helpful, charismatic and always wears a smile. I would recommend Ashutosh to anyone and everyone.
Hardworking and sincere. Extremely reliable and most of the critical tasks in the project were executed by Ashutosh. Technically a strong individual and always helping his team. Would like to work with him again in future.
Ashutosh was highly valued team member while I acted as onsite team lead for couple of assignments for a US client. I found him very hard working and technically sound and a great team player. He is a quick learner and at the same time outperforming with his existing skills. Wish him all the best for his future assignments!
Ashutosh is a focused and committed individual. He ensures that the deadlines are never missed and quality is not compromised. He is an excellent team player, and understands the client business problems quite well.
Ashutosh is very dedicated and hard working person with very good technical skills. He has helped other colleagues in office, showing his sincerity and generosity.
Ashutosh has a very good capability of resolving all critical issues faced by the team of 10 developers. He has always kept work on priority and ensured successful completion with dedication. He is technically very sound and an excellent team player. I wish him good luck.
M.S. Software Engineering (cloud computing and enterprise software), San José State University, 2016–2018. GPA 3.74, Phi Kappa Phi, ACM
B.E. Computer Engineering, Don Bosco Institute of Technology, University of Mumbai, 2007–2011. IBM technical contest prize winner
Keeping up: I read new research through Paper Desk, one of the agents above.
09 / 09
Let's build AI that holds up in production
I'm in Santa Clara, CA, working on enterprise AI: retrieval, agent workflows, Azure architecture and the full-stack apps around them. The quickest way to reach me is email.