Senior Software Engineer, Applied AI · Bengaluru

I build the control plane for enterprise AI.

Gateways that every model call has to pass through. Policy enforced fail-closed at runtime. Agents with identities of their own, and an audit trail that still holds up when there is nobody left to ask. Four years of shipping production platforms end to end: at a health-tech startup, in BFSI at Mphasis, and in Praxis, a platform I built and launched on my own.

See Praxis Read the writing Download résumé

500+

agents in production on the platform I architected

1M+

agent requests a day, every one policy-checked

10M+

messages a day through a status service I built

4M+

medical images classified every month

Independent product

Designed, built and launched solo.

Praxis

The platform for your AI workforce.

Visit live product

Praxis builds agents from a versioned spec, proves they're ready (red-team tested, evaluated, policy-bound), and runs them inside existing business processes next to the human team, with one audit trail across every step.

Golang · TypeScript · AgentScope · LiteLLM · Temporal · OPA · Keycloak

i.Identity-aware gateway

Every LLM and tool call passes one chokepoint that verifies identity, enforces role-based authorization, runs inline prompt-injection and jailbreak guardrails, and signs an audit event. Agents are non-human identities minting short-lived tokens. No standing keys, and suspension revokes issuance instantly. Joiner, mover and leaver lifecycle, access reviews and SCIM 2.0.

ii.Attribution & audit

Token spend and guardrail events attributed to the exact human or agent. Provider-agnostic routing with budgets. Humans, agents and service calls land on one signed, tamper-evident, object-locked timeline you can query for who did what, under which identity, against which model and data.

iii.Proof before production

The WARDEN red-team gate runs adversarial and functional tests mapped to the OWASP LLM and Agentic Top 10, MITRE ATT&CK, GDPR and the EU AI Act, and can block a deploy. Agent Evals grade runs on versioned datasets. A Policy Engine compiles plain-English rules to Rego, enforced fail-closed by OPA.

iv.Build & run

Agents are declared as a spec plus a role file, generated onto AgentScope, and each runs as its own service. Risky tools pause for human approval. A Workflow Composer runs business processes as durable Temporal graphs, with a human gate on every state-changing call. Driven by a CLI and a builder skill for Claude Code and Cursor.

Work

Platforms that other teams, and other agents, depend on.

Mphasis

Feb 2026 to present

Senior Software Engineer, Applied AI · Bengaluru

ATOM enterprise agentic platform

Architected and shipped a multi-tenant platform for BFSI clients: TypeScript, Golang and FastAPI services, a web Studio UI, and a policy gateway with auth, OPA, rate limiting and an immutable audit log. It now runs 500+ agents at 1M+ requests a day. I also built AEGIS, its security layer, which red-teams agents for prompt injection, jailbreaks and goal hijacking, and enforces PCI-DSS, GDPR and DORA guardrails.

TypeScriptGolangFastAPIReactKubernetesKafka / RedpandaOPAPostgreSQLRedis

OathKeeper intelligent RPA for BFSI operations In progress

Today you can automate an SOP with RPA, but the intelligence is lost: the factors that made a BPS operator take a decision never make it into the bot. OathKeeper keeps that oath. It captures operators working across Java, C#, web, mail and mainframe applications, with redaction applied at capture and every step audited, and moves each process through four stages until the automation runs it end to end. The recorded behaviour, screenshots and sessions then train a foundation model to decide the next action itself, cutting human-in-the-loop down to the decisions that truly need a person.

RecordReplayShadowTakeover

RPAJavaC# / .NETWebMailMainframeFoundation modelsPII redactionAudit

Ontologic an AI co-designed knowledge graph

Domain experts and an AI co-designer build an ontology together. The AI proposes classes and relations, a named human confirms each one, and the approved model ships as validated Turtle onto ArcadeDB, with Apache Jena as the reasoning layer. Bi-temporal provenance, append-only corrections and a single validated write path mean the system can say what it believed at any past moment, what is missing, and where sources disagree.

RDF / OWLTurtleSHACLApache JenaArcadeDBLLMs

IHX a Perfios company

Aug 2022 to Feb 2026

Senior Software Engineer, Jul 2024 to Feb 2026 · Software Engineer, Aug 2022 to Jul 2024

Document Intelligence Platform

The document-processing platform for India's largest health benefits administrator. It lifted first-instance complete claim submissions from 23% to 77% within two weeks, and now classifies 4M+ medical images a month.

TypeScriptFastAPIAWS EKSMongoDBS3

Persistence Engine

A single-layer persistence solution that cut average query time from 1–2 s to 10–15 ms across ten tables.

TypeScriptTypeORMPostgreSQLRedis

Status Notification Service

A high-throughput system processing 10M+ messages a day to track state across AI/ML entities, giving clients one real-time view of status.

TypeScriptAWS SQS / SNSMongoDB

Workflow Orchestration Engine

Orchestration for document pipelines (extraction, OCR, classification and QC) with parallel processing that cut processing time by 40%.

Netflix ConductorZookeeperRedis SentinelTypeScript

Platform infrastructure & identity

As an early engineer I owned infrastructure and access control end to end: Kubernetes and Helm scaling and hardening for AI-model serving, the full observability stack, and a Keycloak MFA and SSO solution with custom plugins that cut login time by 50%.

KubernetesHelmNGINXAWSGrafanaLokiPrometheusKeycloakOAuth 2.0

Writing

Essays on provenance, time and identity: the parts of a system you only miss once they're gone.

Aug 2026Knowledge graphs · Security

Your Database Has One Clock. It Needs Two.

“Was that host vulnerable on the tenth of March?” has two correct answers. Bi-temporal modeling for security and audit data, and why dwell time is the gap between when something was true and when you learned it.

Aug 2026AI governance · Provenance

Nobody Left to Ask

Automation doesn't make your audit trail less important. It makes it the only thing left. Why provenance becomes the only witness once decisions are automated, and three questions vector and RAG stacks structurally cannot answer.

Mar 2025Keycloak series · Part 3

Building a Custom MFA Plugin for Keycloak: SMS, Email & Telegram

Writing a Java SPI extension that adds SMS, email and Telegram as second factors.

Jul 2024Keycloak series · Part 2

Mastering Multi-Factor Authentication in Keycloak

An end-to-end guide to SMS, email and TOTP flows, from realm configuration to custom authentication flows.

May 2024Keycloak series · Part 1

Orchestrating Centralized User Management with Keycloak

One Keycloak realm in front of many services, with path-level access carried in the token itself.

All posts on Medium

Side projects

Open source, experiments, and the occasional game.

npm package · 500+ weekly downloads

Python Spawn

A bridge between Node.js and Python, with automatic environment detection, robust error handling, and function calls across runtimes.

Node.jsPythonnpm

Real-time · built by hand

Collaborative Compiler

A multi-user code editor with live sync and autocomplete, written before AI coding assistants existed.

JavaScriptWebSockets

Security demo

Secure Files

A full-stack MFA demo on Keycloak: a custom Java SPI plugin, self-signed certificates, RBAC, and SSO across services.

DjangoReactKeycloakJava

Vibe Jam 2025 · built in 2 hours

Contra

A complete 3D run-and-gun shooter with enemy AI, vehicle physics and particle systems, built in two hours with AI tooling. Because why not.

Three.jsAI-assisted

Toolkit

What I reach for, grouped by the problem it solves.

AI governance & security

LLM gateway design, OPA / Rego policy-as-code, agent identity and short-lived service tokens, SCIM 2.0, access reviews, red-teaming (OWASP LLM and Agentic Top 10, MITRE ATT&CK), PII-aware routing, budget governance, immutable audit logging

Agents & LLMs

Multi-agent orchestration, agent evals and grading, human-in-the-loop workflows, intelligent RPA (record, replay, shadow, takeover), RAG, LiteLLM, LangGraph, LangChain, AgentScope

Knowledge graphs

RDF / OWL ontologies, Turtle, SHACL, SPARQL, Apache Jena, ArcadeDB, bi-temporal data modeling

Languages & backend

TypeScript, Golang, Python, SQL · Node.js, FastAPI, Django, Express · REST, WebSockets, Kafka / Redpanda, Temporal, event-driven architecture

Data & infrastructure

PostgreSQL, MongoDB, Redis · AWS (EKS, SQS, SNS, S3, API Gateway, Cognito) · Docker, Kubernetes, Helm, GitHub Actions

Identity & observability

Keycloak, OAuth 2.0, JWT, MFA, SSO · Grafana, Loki, Tempo, Prometheus

About

I started at IHX as an early engineer, processing millions of medical documents a day on systems that weren't allowed to go down. I owned whatever needed owning, from the Kubernetes clusters to the observability stack to the login flow, and learned that the infrastructure around a product is usually where the hard problems live.

That carried straight into AI. The agent logic is often the easy part. Who the agent is, what it's allowed to touch, what it cost, and whether you can prove afterwards what it did: that's where the real engineering is. It's what I build at Mphasis, what I built Praxis to do, and what I keep writing about.

Based in

Bengaluru, India

Education

B.E., Information Science Sapthagiri College of Engineering, 2018 to 2022 · CGPA 8.5

Certifications

Architecting with Google Compute Engine The Modern JavaScript Bootcamp

Off the clock

Away from the keyboard I'm usually headed somewhere new, ideally with a little adrenaline involved. So far that's meant Dubai, Singapore, Bali and Malaysia, and a lot of Indian highway. Closer to home I practise Isha Yoga and volunteer with Isha. It's the stillness that balances everything else.

DubaiSkydiving
SingaporeThe giant swing
Central & South IndiaA 15-day road trip by car
Isha YogaPractitioner and volunteer

Open to conversations

Have a hard problem? Let's talk.

Agentic AI infrastructure, AI security, high-scale systems, or roles that involve real systems at real scale. I reply fast.

shreyasmk.mathur@gmail.com

LinkedIn GitHub Medium +91 88672 29568