Srivatsa Kamballa - Software Engineer

Hi. I'm

A Software Engineer who breaks what he builds.

Tirupati → Bangalore → Chicago
Last Played
On repeat • Spotify •
Fav Tool
Tools
Connect
Email LinkedIn GitHub Instagram
Scratch Me
Hours Coding
0
Coffees Drank
0
Activity
Less
More

srivatsa@chicago: ~/portfolio

$ she asked how much I love her ❤️❓🎺

me: "O(n)."

she asked what that means.

me: "depends on how many problems you give me." 😭

Projects

rag-redteam

rag-redteam | Open-Source RAG Security Tool

Jun 2026 - Present · Open Source

Red-teams retrieval pipelines across 7 vulnerability classes as a CLI and a CI-gating GitHub Action, with deterministic canary detection and zero runtime dependencies. Benchmarked LangChain, LlamaIndex, and Haystack defaults, all exploitable at 50 to 75%, proving injection is a pipeline problem, not a model one.

PythonLLM SecurityGitHub ActionPyPIRAG

kubemend

kubemend | Kubernetes SRE Agent

Aug 2026 - Present · Open Source

A remediation agent that diagnoses freely and acts narrowly: six typed, reversible actions whose blast radius is computable before execution and whose only write surface is a git commit. Now with a published threat model, evaluation, and findings report, a live GitOps demo, and zero runtime dependencies.

KubernetesPythonSREGitOpsAgents

MetARAG

MetARAG | Document Intelligence Platform

Aug 2025 - Dec 2025 · Industry - CCC Intelligent Solutions

Led 5+ engineers on a GPU-accelerated RAG platform covering parsing, chunking, embeddings, and retrieval tuning, reaching about 93% retrieval precision across 100+ GB with about 65% better processing efficiency and 40% lower latency.

RAGLangChainGPUAWSKubernetes

ECI Pipeline

ECI Pipeline | Android Risk Intelligence

Jan 2026 - May 2026 · Industry - TransUnion

A DeltaRAG and Graph-RAG platform on AWS Lambda ingesting 10 live Android security and CVE feeds, turning them into evidence-backed risk tickets with a Sentinel scoring agent and a live Next.js dashboard.

FastAPIpgvectorGraph-RAGNext.jsAWS Lambda

Chain-of-Thought on CLEVR

Chain-of-Thought on CLEVR

Jan 2026 - May 2026 · Research - CS533

Fine-tuned BLIP-2 with LoRA on 50,000 CLEVR samples, lifting accuracy from 8.75% zero-shot to 45.95%. A controlled study found chain-of-thought wins on short reasoning chains and loses on long ones, the opposite of the usual assumption.

PyTorchBLIP-2LoRAVision-Language

End-to-End DevSecOps EKS Platform

End-to-End DevSecOps EKS Platform

2025 · Platform - SRE

A production 3-tier application on AWS EKS with 8 microservices and zero-downtime deploys, automated end to end with Jenkins, ArgoCD, Trivy, and SonarQube gates, provisioned with Terraform and observed with Prometheus and Grafana.

AWS EKSTerraformArgoCDJenkinsDevSecOps

Movie Recommendation

Movie Recommendation | MLOps Lifecycle

The full recommender lifecycle, not just a notebook: training, model serving, A/B tests, drift detection, and Prometheus and Grafana dashboards, reaching RMSE 0.58.

PythonDockerPrometheusGrafanaMLOps

Counterfactual Fact Verification

Counterfactual Fact Verification

2026 - Present · Research

Zero-shot fact-checking on FEVER with local quantized LLMs, generating counterfactual claim variants across complexity tiers to map where small models stay honest and where they break.

PythonFEVERPhi-3Llama 3.1Mistral

Open Source

0

merged

0

open

0K★

combined stars

0

major projects

pydantic/pydantic #13374+9 −0

Mergedapproved by a maintainer

Stopped a JSON schema config being silently dropped

When an Annotated field combined a dict json_schema_extra with a callable one, Pydantic warned the callable would be ignored, then did the opposite: it dropped the dict and could crash schema generation. I made the dict win as documented and added a regression test. Approved by a core maintainer.

BerriAI/litellm #30764+59 −5

Stopped short secrets leaking into logs

SensitiveDataMasker returned any secret of 8 characters or fewer verbatim, leaking short redis passwords, API keys, and tokens into logs and the admin UI. I made short values mask by default, audited all 16 call sites, and added regression tests.

run-llama/llama_index #22133+68 −9

Fixed silent data loss in the ingestion core

In IngestionPipeline upserts, nodes were keyed by document id in a dict, so every chunk of a document except the last was silently dropped, never embedded or stored. I fixed both the sync and async paths with regression tests for each.

BerriAI/litellm #29693+36 −2

Fixed a 10x cost-tracking error

amazon.titan-embed-text-v2 was priced ten times too high in the cost maps, inflating everyone's budget reports. I verified the number against AWS Bedrock pricing and fixed the primary and backup maps in sync so they would not drift apart again.

deepset-ai/haystack #11670+27 −0

Silenced noisy logs on empty inputs

HTMLToDocument dumped ERROR lines on empty ByteStream inputs because lxml logs its parse failures internally rather than raising. I skipped empty streams before extraction, and separately authored the OpenSearch and Elasticsearch SQL retriever docs (#11543, #11494).

BerriAI/litellm #31725+259 −8

Time-based off-peak pricing in the cost engine

A feature on a core billing path: models define a UTC window and discounted rate, and the calculator picks the rate by time of day, backward-compatible, with regression tests. Open and under review.

stanfordnlp/dspy #9942+24 −2

Fixed value corruption in the chat parser

In ChatAdapter.parse, the field-header regex matched the stripped line but sliced the unstripped one, leaking stray marker characters into parsed values on indented headers. I aligned the match and the slice with regression tests. Open and under review.

Writing

kubemend: failing safe and failing silently are the same mechanism.

Failing safe and failing silently are the same mechanism

A component I built so it could never break the main path turned out to be able to hide its own bugs. The two properties arrive together, and separating them took a per-request connection and a real socket.

rag-redteam cover: injection is a pipeline problem, not a model one.

RAG injection is a pipeline problem, not a model problem

I built rag-redteam to test the pipeline that turns untrusted documents into trusted context. A frontier-model upgrade did not fix injection. That is the whole point.

When to stop fuzzing: a Bayesian answer to a wall-clock question.

When is it actually safe to stop fuzzing?

Most fuzzing campaigns stop on a wall-clock budget and a shrug. PSBayes estimates the residual risk you are actually carrying, per seed.

Also onMedium Dev.to

Experience

Software Developer, Software Testing • University of Illinois Chicago

Aug 2026 - Present • Chicago, IL

First author on PSBayes, a per-seed Bayesian estimator of residual fuzzing risk. Paper under submission.

Software Developer • University of Illinois Chicago

Aug 2024 - Jul 2026 • Chicago, IL

Rebuilt FuzzBench on AWS (~60% faster setup), provisioned 10+ nodes with Terraform and Ansible, cut MTTR from 3 hours to under 1.

Software Developer, AI Engineer • UIC College of Business Administration

Aug 2025 - May 2026 • Chicago, IL

Led 5+ engineers on MetARAG, a GPU-accelerated RAG platform reaching ~93% retrieval precision, and shipped the ECI Pipeline for TransUnion.

Trainee Software Engineer • Mu Sigma

May 2023 - Jun 2024 • Bangalore, India

Automated Python and SQL ETL for enterprise decision-support clients, cutting recurring data-prep effort ~40%.

Education

Master of Science, Computer Science • University of Illinois Chicago

Aug 2024 - May 2026 • Chicago, IL

GPA 3.88 / 4.00

B.Tech, Computer Science (AI Specialization) • Jain University

2020 - 2024 • Bangalore, India

GPA 3.91 / 4.00

Certifications, Patent & Research

AWS Certified Solutions Architect – Associate

Amazon Web Services · Issued 2026

Designing distributed, cost-optimized, and resilient systems on AWS: VPC and IAM, auto-scaling, and observability.

Book Issue Management System for Libraries

Patent No. 202341071153 · Intellectual Property India · Nov 2023

Co-invented an AI-enabled camera and RFID system for automated library book issuance, return, and access control.

PSBayes: Bayesian Residual-Risk for Fuzzing

First-authored paper · under submission

A per-seed Beta-posterior estimator of the undiscovered risk left in a fuzzing campaign, evaluated across 8 FuzzBench benchmarks and 18,861 discovery rounds.