Skip to content
View Srivatsa03's full-sized avatar

Highlights

  • Pro

Block or report Srivatsa03

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Srivatsa03/README.md

~/ whoami

$ cat about.txt

Hi, I'm Srivatsa Kamballa. I build the unglamorous half of software: the pipelines, the infrastructure, and the AI systems that have to keep working after the demo crowd goes home. Most of what I do lives in the gap between "it runs on my machine" and "it survives production at 3am."

  • Currently building rag-redteam, mcp-snitch (session-level MCP attack detection), and minja-repro (reimplementing a no-code LLM-agent attack paper)
  • Portfolio: srivatsa-kamballa.vercel.app
  • MS Computer Science @ University of Illinois Chicago (GPA 3.88)
  • Open to full-time Software, Platform / Infrastructure, and AI Engineering roles across the US
  • Fun fact: I fix the bugs everyone else scrolled past in libraries you already import (10 merged PRs across Pydantic, LiteLLM, LlamaIndex, Haystack)

~/ toolbox

tech stack

~/ language radar

language radar chart

~/ contribution calendar

3D isometric contribution calendar



snake eating the contribution graph

~/ the numbers

GitHub statistics

~/ selected work

rag-redteam minja-repro
kubemend ECI-Pipeline
Chain-of-Thought-on-CLEVR Counterfactual_Fact_Checking
Movie-Recommendation DevSecOps on AWS EKS
WuzzyFuzz-03 Telegram-MTproto2.0

Each one has a full write-up (how it works, the tradeoffs, what broke, and measured results) at srivatsa-kamballa.vercel.app/projects.


~/ open source

Those repos that look like forks of LiteLLM, LlamaIndex, Pydantic, and Haystack? That's where I actually fixed things. 10 merged, 4 open so far.

  • LiteLLM (57k star): shipped time-based off-peak pricing in the cost engine, a backward-compatible feature on a core billing path. Caught an equal-ended window silently billing standard rates around the clock, and escalated the schema to maintainers rather than guess. PR #31725, +1,039/-25, merged.
  • LiteLLM (57k star): fixed double-billing when cached and image tokens overlapped on multimodal requests. PR #37407, merged.
  • LlamaIndex (51k star): silent data loss where all but the last chunk of a document was dropped on upsert. PR #22133, merged.
  • Pydantic (28k star): a json_schema_extra dict was silently dropped, and schema generation could crash, when a callable followed it in an Annotated type. PR #13374, merged, maintainer-approved.
  • LiteLLM (57k star): a masker leaked short secrets (8 chars or fewer) into logs and the admin UI. PR #30764, merged.
  • Haystack (26k star): silenced noisy ERROR logs on empty inputs, plus authored the OpenSearch and Elasticsearch SQL-retriever docs. PR #11670, #11543, #11494, merged.

Plus the 10x embedding-pricing fix (#29693), a None-metadata crash fix in LlamaIndex (#22773), and 4 open PRs across LiteLLM, DSPy, and instructor (an SSRF fix that rejects IPv6 transition addresses embedding private IPv4).


~/ one patent, for the curious

Book Issue Management System for Libraries (IP India, granted Nov 2023, co-invented): AI cameras, RFID, and automated access control for hands-off library check-in and return.



$ logout  ·  built to survive 3am  ·  thanks for scrolling

Pinned Loading

  1. Chain-of-Thought-on-CLEVR Chain-of-Thought-on-CLEVR Public

    Fine-tuned BLIP-2 + LoRA on CLEVR to map where chain-of-thought helps visual reasoning and where it hurts. 8.75% -> 45.95%.

    Python

  2. E2E-K8s-Three-Tier-DevSecOps-Project E2E-K8s-Three-Tier-DevSecOps-Project Public archive

    JavaScript

  3. ECI-Pipeline ECI-Pipeline Public

    DeltaRAG + Graph-RAG pipeline turning 10 live Android security feeds into evidence-backed risk tickets. 93% retrieval precision.

    Python

  4. fuzzbench fuzzbench Public

    Reproducible fuzzing research environment (Google FuzzBench + custom LLVM build) with Bayesian residual-risk estimators.

    Python

  5. Movie-Recommendation Movie-Recommendation Public

    End-to-end MLOps: training, serving, A/B tests, drift detection, Prometheus/Grafana. RMSE 0.58.

    HTML 2

  6. My-Portfolio My-Portfolio Public

    Portfolio of Srivatsa Kamballa

    TypeScript 1