I build systems.Then test wherethey break.

I’m Vipul Ponugoti, a software engineer building reliable backends and distributed systems. SDE intern at VibeLevel.ai · B.Tech CSE (AI & ML), ’27 · Vadodara, India.

Selected work Resume (PDF)

Live model: one part fails, traffic reroutes, it recovers.

876open-source pull requests merged 0writes lost in 50 crash tests 250korders matched per second 2engineering internships

01Selected engineering projects

Systems built forthe difficult parts.

Three projects exploring consensus, low-latency execution, and reliable AI infrastructure. Inspect the architecture, measurements, and failure cases.

01 / Distributed systems · Consensus · Fault tolerance

Raft-replicated distributed key–value store

A five-node key–value store implementing Raft from scratch in Java 21, built to examine consistent replicated state through crashes, partitions, delayed messages, and leadership changes.

Java 21gRPCDockerJUnitJMH

Architecture & failure testing

Leader election and replicated log coordination across five nodes
gRPC key–value API designed for linearizable operations
Log replication and committed-operation propagation
Snapshotting for replicated-state management and recovery
Fault-injection harness for crashes, partitions, and message delays

Source not publicProject details in résumé

RAFT / FIVE-NODE CLUSTERARCHITECTURE STUDY

LEADER REPLICATING ENTRIES

replicated nodes

fault-injection scenarios

lost acknowledged writes

20k

operations / second

12 ms

reported p99 latency

350 ms

measured failover

Project benchmark: five-node cluster, 50 fault-injection scenarios. Throughput and latency describe the tested workload.

02 / High-performance systems · Event-driven architecture

Low-latency exchange matching engine

An exchange-style order matching engine focused on deterministic execution, price–time priority, low-latency processing, and real-time market-data distribution.

Java 21Spring BootApache KafkaRedisWebSocketJMH

Architecture & failure testing

Price–time-priority order book for limit and market orders
Kafka event recording for deterministic replay of order processing
WebSocket-based market-data fan-out
Redis-backed per-client rate limiting
Custom load generation and JMH-based performance evaluation

Source not publicProject details in résumé

MATCHING / DETERMINISTIC FLOWARCHITECTURE STUDY

LIMIT / MARKETPRICE → TIME

ORDER APIVALIDATE / RATE LIMIT

ORDER BOOKMATCH / EXECUTE

KAFKA LOGAPPEND / SEQUENCE

REPLAYDETERMINISTIC STATE

MARKET DATAWEBSOCKET FAN-OUT →

Price–time priority · event sourcing · custom load generation

250k

orders / second

50 μs

reported custom-benchmark p99

10k

concurrent WebSocket clients

Custom load-generator benchmark. The 50 μs p99 measures matching, not full end-to-end network latency.

03 / AI infrastructure · Cloud-native engineering · Reliability

Multi-provider LLM gateway on Kubernetes

A Kubernetes-deployed gateway for provider availability, response latency, usage cost, tenant governance, and observability around applications integrating large language models.

JavaSpring BootSpring AIKubernetesPostgreSQLpgvectorRedisGitHub ActionsOpenTelemetryPrometheusGrafana

Architecture & failure testing

Multi-provider request routing and provider failover
Per-tenant token budgets and Redis rate limiting
PostgreSQL + pgvector similarity-based semantic caching
Kubernetes deployment with autoscaling and GitHub Actions CI/CD
OpenTelemetry, Prometheus, and Grafana observability
Resilience testing during simulated LLM provider outages

Source not publicProject details in résumé

LLM GATEWAY / REQUEST PATHARCHITECTURE STUDY

APPLICATION

TENANT REQUEST

LLM GATEWAY

RATE LIMIT / BUDGET
ROUTE / OBSERVE

PROVIDER ROUTE

FAILOVER ROUTE

PGVECTOR

SEMANTIC CACHE

REDIS

RATE / BUDGET

KUBERNETES

SCALE / DEPLOY

OTEL TRACESPROMETHEUS METRICSGRAFANA

Provider failover · tenant budgets · semantic cache · OpenTelemetry

37%

lower average latency

36%

lower token cost

40%

semantic-cache hit rate

Tested

simulated provider outages

Project benchmark with semantic caching and simulated provider outages. Results depend on the model, provider, and workload.

02Professional experience

Production work.End-to-end responsibility.

Internship contributions across AI-native evaluation software and healthcare-oriented machine learning, spanning backend services, frontend delivery, reliability, testing, and model evaluation.

AI-native skills assessment and hiring technology

Backend development and full-stack feature delivery for software evaluating technical capabilities and human–AI collaboration, with an emphasis on evaluation correctness, testing, reliability, and performance.

PythonFastAPITypeScriptReactNext.jspytestMCP

Documented responsibilities4 work areas

Designed and developed an LLM-based scoring engine for technical sessions. A central challenge was correcting attribution errors where AI-performed work could be assigned to the human participant, or vice versa. The implementation is supported by approximately 2,700 lines of pytest coverage.

03How I work

Understand the system.Test its limits.

Java and Spring are my foundation, with Python and TypeScript for APIs, AI infrastructure, and full-stack work. I’m looking for software engineering roles where reliability and performance matter.

Trace the failure

Understand existing behavior and reproduce the edge case before changing code.

Measure the trade-off

Use benchmarks and observability to guide performance and design decisions.

Keep the regression test

Make the fix reproducible and leave the next engineer a way to verify it.

04Open-source engineering

Read the system.Improve the behavior.

Contributions centered on application security, concurrency, responsiveness, automated testing, and maintainability across unfamiliar codebases and established project architectures.

876

Merged pull requests

projects contributed to

351

merged in Truxify

204

merged in ai-code-reviewer

linked examples below

Contributed inNode.jsPython / FastAPIRustDart / FlutterSolidity

As listed on my resume Browse merged pull requests

Selected contribution ledger

TruxifySecurity

Raw-body webhook verification

Implemented HMAC-based webhook signature verification using the original raw HTTP request body, avoiding validation against transformed payloads.

HMAC · WebhooksMerged #5296

TruxifySecurity

Path traversal protection

Improved path-handling security to protect against unauthorized file access through traversal patterns.

Filesystem securityMerged #5313

TruxifyTesting

Vitest testing foundation

Added the Vitest framework and pricing tests to make backend behavior repeatable and catch regressions. Rated exceptional by the maintainers.

Vitest · Regression testingMerged #303

ai-code-reviewerSecurity

Symlink race-condition hardening

Addressed a symlink-related time-of-check/time-of-use race condition using O_NOFOLLOW.

O_NOFOLLOW · TOCTOUMerged #3470

ai-code-reviewerConcurrency

Atomic circuit-breaker admission

Implemented atomic admission handling to improve circuit-breaker consistency under concurrent access.

Concurrency · Circuit breakersMerged #3464

ai-code-reviewerPerformance

Incremental Redis key scanning

Replaced potentially blocking Redis KEYS operations with incremental SCAN operations to improve responsiveness.

Redis · SCANMerged #3472

peer-learningSecurity

Timing-safe JWT comparison

Replaced JWT signature string comparison with timingSafeEqual, with length checks and regression tests.

JWT · Timing-safe comparisonMerged #681

05Technical toolkit

Tools in context.Fundamentals underneath.

A primary foundation in Java and the Spring ecosystem, complemented by Python, TypeScript, modern frontend development, data infrastructure, and cloud-native operations.

015 areas