Ashutosh Siddh

Senior Engineer/Applied AI

I build AI that ships.

Eleven years leading .NET and AWS delivery across insurance, telecom, healthcare and automotive. I build LLM products the same way: correct, permissioned, cheap to run.

Ashutosh Siddh
Based
Gurgaon, India (UTC+5:30)
Open to
Remote, worldwide
Shipping since
2015
Agent orchestrationRetrievalModel Context ProtocolEvaluation harnessesGuardrailsFailure taxonomiesVoice agentsMulti-tenant dataAdversarial testingObservabilityAgent orchestrationRetrievalModel Context ProtocolEvaluation harnessesGuardrailsFailure taxonomiesVoice agentsMulti-tenant dataAdversarial testingObservability

How I got here

2015 to 2024

Enterprise delivery

Eleven years of .NET, C# and SQL Server across insurance, telecom, healthcare and automotive. Led delivery of an insurance platform, modernised a telecom system off legacy VB and Oracle, and ran teams alongside the code. Regulated domains, real consequences.

The turn

Models changed the problem

A system that returns a confident, well-formed, wrong answer passes every check an enterprise team already owns. That is an engineering problem before a research one, and it is the one I had been solving for a decade: is this right, who is allowed, what does it cost.

Now

Same discipline, new material

Agents, retrieval, governed tool access, evals, observability. Built the way anything that has to survive production is built: deterministic where it can be, gated where it must be, tested against the failure rather than the happy path.

What that produces

Five systems, built end to end. Each one ships with a written account of what it refuses to do.

Anvaya Platform

Multi-tenant control plane for an AI security product

A test that has never failed constrains nothing. An isolation suite written against the schema it validates will pass on the first run and tell you nothing about whether your tenants are actually separated.

Read the case study

before actionAPI keyscoped, not roledSessionArgon2id · TOTPRouterpermission gateTenantScopetype layerwithTenantSET LOCAL · app roleRLS forcedevery org_id tableAudit loginsert-only grant

TENANCY BOUNDARY

Anvaya

AI failure observability and causal diagnosis

An AI product's worst failures return HTTP 200, in perfectly valid JSON, at normal latency, and every classical monitor stays green.

Read the case study

spansincidentYour AI app@anvaya/sdkOTel GenAIno SDK neededIngestspans → tracesL0-L2structural · statL3 judgeopt-in · billedAttributionspan-tree causeTriage56-mode codes

DIAGNOSIS PIPELINE

ConciergeAI

AI customer service the model cannot over-promise on

If the model can issue the refund, then one prompt injection away, it will. So the model never holds the authority to commit the business.

Read the case study

contextproposesdecisionCustomerchat widgetRetrievalBM25 · verbatimRails inPAN redactionLLMproposesPolicydecides · no LLMRails outcommitmentLLMnarrates

AUTHORITY BOUNDARY

AegisQuery

Governed analytics MCP server

A read-only transaction is not a security boundary. Proving that took reproducing the exact injection that defeats one.

Read the case study

tools/callallowedLLM agentMCP clientMCP serverstdio · HTTPGovernanceallowlist · PIIAST guardsingle SELECTCost guardplan budgetWarehouseread-onlyAuditper call

SAFETY PIPELINE

Dialoft

AI voice SDR for outbound sales and qualification

Phone and browser voice are different transports, not different products. One conversation brain serves both, or the two drift apart within a month.

Read the case study

messagesreasontool callsVapiphone channelLiveKitweb channelDeepgramNova/Aura STT+TTS/v1 endpointOpenAI-compatibleLangGraph8-node graphClaudereasoning8 toolsCRM, KB, DNCLead scoreBANT 0-100

VOICE TOPOLOGY

Virtual Buddy

AI-employee platform for creators

An unattended job that silently stops firing looks exactly like a healthy one. Alerting cannot cover that, because an alert only fires from a job that ran.

Read the case study

streamDrizzleWeb appNext.js 16Mobile appExpo / RNVercel cronbriefingsAPI routesClerk authClaude agenthaiku → opus tiersToolstrends / dealsNeon PostgresRLS · 14 tables

AGENT LOOP

Also built

  • evalharnessLLM evals with a CI quality gate
  • KESSLEROpen orbital conjunction assessment toolkit
  • PENUMBRAProbabilistic space-weather forecasting
  • RAG PlaygroundA modular retrieval lab
  • Virtual Buddy MobileiOS and Android companion
  • Sir Roasts-A-LotPersona and prompt engineering, shipped
  • SMSERPSchool management ERP

Most AI incidents are not model failures. They are authority failures.

A model that can move money will eventually move it for the wrong reason. A query tool with a read-only connection will eventually meet the injection that walks around it. The interesting engineering is not in the prompt. It is in deciding what the model is structurally unable to do, and then proving it in a test suite.

The eleven years in full

Six companies, five domains, writing the code through to owning the delivery.

Apr 2025 to Present

TrueBlue

Software Engineer

Build and maintain a workforce-management product: backend services in .NET Core and Web API over SQL Server, deployed on AWS. Claude and GitHub Copilot are part of the daily development and review workflow.

Jun 2021 to Mar 2025

GlobalLogic

Associate Consultant

Led end-to-end delivery of an enterprise insurance platform (C#, ASP.NET Web API, MVC, .NET Core, SQL Server), from requirements through unit testing, with combined development and team-lead responsibilities.

Sep 2019 to Jun 2021

Nagarro

Senior Associate

Built a customer and dealership platform (.NET, Oracle, AWS Lambda) with Terraform and TeamCity CI/CD, and delivered an HR staffing application as microservices with an Angular front end.

Oct 2017 to Aug 2019

LiquidHub (Capgemini)

Technical Lead

Designed and built in-house Google and Microsoft authorization API integrations end to end using .NET (C#, Web API) and SQL Server.

Jun 2017 to Oct 2017

GGK Tech

Software Engineer

Developed healthcare web products for one of the largest US health-benefits networks using C#, MVC, Web API, and Angular.

Jul 2015 to Mar 2017

Infosys

System Engineer

Modernized a large telecom system for a major Australian telecom provider: migrated legacy VB to C# and Oracle to SQL Server, and built data-migration and automation-test tooling.

About

I spent eleven years leading and delivering enterprise .NET platforms across insurance, automotive, telecom, and healthcare. Now I build production LLM systems end to end: agent orchestration, retrieval, voice, governed tool access over the Model Context Protocol, evaluation harnesses, and AI-specific observability. The through-line is making non-deterministic systems safe to operate. Deterministic policy engines the model cannot bypass, adversarial test corpora, CI gates that fail the build on regression, and failure taxonomies that say why a system broke. Everything ships: AWS (EC2, Lightsail, Lambda), Vercel for frontends, Railway and Render for backends.

Stack

AI engineering
Claude API, Agent tool-calling, Multi-agent systems, LangGraph, LangChain, RAG pipelines, MCP servers, Voice agents, LLM-as-judge evals, AI observability, Guardrails and red-teaming, Prompt-injection detection, Structured outputs
Security
Multi-tenant isolation, Row-level security, Threat modelling, Adversarial and mutation testing, RBAC and permission gates, Argon2id, sessions, TOTP MFA, Append-only audit logs, OWASP LLM Top 10
Backend
.NET Core, C#, ASP.NET Web API, Microservices, Python, FastAPI, TypeScript, Node.js, Next.js
Data
SQL Server, Oracle, Postgres / Neon, pgvector, SQLite, Drizzle, Prisma, PGlite, Schema migrations
AWS
EC2, Lightsail, Lambda, S3, RDS, VPC, IAM, CloudWatch, ECS / ECR, CloudFormation
Platform
Docker, Nginx, Terraform, GitHub Actions, TeamCity, Azure DevOps, Vercel, Railway, Render

Education

  • M.Sc. in Information Technology

    DA-IICT, Gandhinagar

    2013 to 2015

  • Bachelor of Computer Applications

    Ahmedabad University

    2010 to 2013