Tool-calling agents
Multi-step agents with typed tools and MCP servers, wired into real APIs and databases.
Open for collaboration
Systems that don’t drop events. Agents that don’t drop the ball.
I’m Saurav Modi, a Senior Software Engineer with 10+ years building event-driven, cloud-native platforms for fintech, insurance and higher education. Now I’m bringing that reliability discipline to AI agents: tool-calling, evals, and durable workflows that hold up in production.
agent-run · notice-triage
eventnotice.created · p3 · offset 1,048,213
planclassify → enrich → draft → verify3 steps
toolpostgres.query(account_id)ok 12ms
toolkafka.replay(dlq, n=42)ok 38ms
evalpolicy_check · pii_scanpass
retryidempotency-key reused · no dup1×
done1 message queued · 0 lost 1.4s
illustrative trace
10+years of experience
1M+Kafka events per day, zero data loss
99.99%availability after ECS migration
−40%Aurora PostgreSQL cost
Now · Agents & AI
An agent is a distributed system that happens to think.
LLM agents fail the way microservices fail: timeouts, partial writes, retries that double-send, silent drift. I design agentic systems with the guardrails I’ve spent a decade building into event pipelines.
Durable agent workflows
Agent steps run as events on queues, with checkpoints, idempotent tools and human approval where it matters.
Evals & guardrails
Golden datasets, LLM-as-judge scoring, PII and policy checks in CI, so prompt changes ship like code changes.
RAG on enterprise data
Retrieval over Postgres and pgvector with access control, freshness and citations.
LLM observability
Traces per run, token and cost budgets, latency targets, and replayable runs for debugging bad answers.
AI-assisted engineering
Coding agents in the daily loop for reviews, migrations and tests, with humans owning the merge.
What Kafka taught me about agents
The patterns carry over almost one to one.
| Event platform | Agent system | Why it matters |
|---|---|---|
| Dead-letter queue | Failed-run queue | Bad tool calls get parked and reviewed, not lost or looped. |
| Idempotency key | Safe tool retries | A retried “send notice” never sends twice. |
| Event replay | Trace replay for evals | Yesterday’s failures become tomorrow’s regression tests. |
| Backpressure | Token & rate budgets | One runaway agent can’t exhaust shared model quota. |
| Exponential backoff | Model fallback chain | Provider hiccups degrade gracefully to a smaller model. |
Experience
A decade across fintech, insurance and higher education.
Building scalable systems and leading engineering teams.
08/2024 – Present
Senior Software Engineer
Virtusa · JPMorgan Chase
- Architected a horizontally scalable, event-driven microservices platform on Apache Kafka handling 1M+ events/day.
- Built automated retries with exponential backoff and DLQs. Designed RESTful APIs with OpenAPI 3.0.
- Migrated 5+ microservices to AWS ECS at 99.99% availability.
- Optimized AWS Aurora PostgreSQL, cutting costs by 40%. Ran chaos engineering exercises to find single points of failure.
Apache KafkaAWSGraphQLPostgreSQLNode.jsDockerKubernetes
08/2022 – 03/2024
Software Engineer
DePaul University
- Optimized the student information system to support 500+ concurrent users.
- Built an API gateway on Spring Cloud Gateway with rate limiting and JWT validation.
- Automated deployment pipelines, cutting deployment time by 40%. Architected React component libraries for UI consistency.
Spring BootReactAWSJenkinsGitHub ActionsTypeScript
11/2016 – 03/2022Switzerland / India
Senior Software Engineer
Tata Consultancy Services · Generali
- Led a monolith-to-microservices migration, improving deployment frequency by 30%.
- Integrated Redis caching, cutting response times by 50%.
- Led an Oracle PL/SQL ETL migration of 5M+ records. Orchestrated CI/CD with Jenkins and SonarQube.
JavaMicroservicesRedisOracleJenkinsSpring Boot
Selected work
Platforms I’ve architected and shipped.
1M+events / day
Event-driven microservices platform
Kafka cluster with deliberate partition-key design, DLQs and a Swagger-based event replay utility. Zero data loss on critical workflows for JPMorgan Chase communications.
Apache KafkaNode.jsPostgreSQL
−30%development cycles
GraphQL data distribution layer
A federated GraphQL layer for declarative data fetching across product teams, improving API consistency.
GraphQLNode.jsRedis
99.99%uptime
Multi-region AWS ECS migration
Active-passive disaster recovery with automated failover, and 25% lower infrastructure cost.
AWS ECSDockerCloudFormation
40+components
React component library
A Tailwind-based design system documented in Storybook, cutting component development time by 50%.
ReactTypeScriptStorybook
Toolkit
What I reach for.
AI & agents
- LLM APIs & tool use
- Model Context Protocol
- RAG · pgvector
- Evals & guardrails
- Coding agents
Languages
- Java 8 / 11 / 17 / 21
- TypeScript
- JavaScript (ES6+)
- SQL · NoSQL
Frameworks
- Spring Boot
- Spring Security
- Node.js
- React · Angular
- GraphQL · Hibernate
Cloud & infra
- AWS ECS · Lambda
- S3 · RDS · SQS
- Docker · Kubernetes
- Jenkins
- GitHub / GitLab CI
Data
- PostgreSQL · Aurora
- MySQL · Oracle
- MongoDB
- Redis
Building something with agents? Let’s talk.
Have a project in mind or want to collaborate? I’d love to hear from you.