// AI Platform & Agentic Systems

ARVIND KUMAR

|

I build production-grade AI agent platforms — from security scanning systems to autonomous multi-agent pipelines. Currently deep in LangGraph, FastAPI, and everything that makes agents actually reliable. Dual-degree: RKGIT B.Tech IT · IIT Madras BS Data Science · WorldQuant Research Consultant.

Top 20% · International Quant Championship 2025RKGIT B.Tech IT · IIT Madras BS Data ScienceOpen to internships & hackathons
Arvind
// Explore

The Neural Graph

Every project and every tool I use, in one connected map. Click a project to isolate exactly the tech stack behind it.

Click a project square (or a legend chip) to isolate its stack · click again to clear
// Projects

Selected Work

// Achievements

Wins & Recognition

TOP 20%
Intl Quant Championship 2025
Global quantitative research competition
GOLD LEVEL
WorldQuant BRAIN
Research Consultant — designing & testing alphas
COMPLETED
Summer Analytics
Applied data science program
FINALIST
Strategy Storm — IIT Guwahati
Global Finalist, Case Competition
FINALIST
Investomania — IIT Hyderabad
Case Competition Finalist
TOP 10
Case Crackers — Delhi University
Among top 10 Nationally
1st RUNNER-UP
Hackwarts — RKGIT
2nd position, team LazyCoders
// Snapshot

Stack & Tools

Core stack

Python
FastAPI
Next.js
MongoDB
LangGraph
Groq

Tools

LiteLLM
ChromaDB
PostgreSQL
XGBoost
Docker
NetworkX
Socket.IO
JWT + bcrypt
Razorpay
SQLAlchemy + Alembic
Motor
HuggingFace
BeautifulSoup + PyMuPDF
Faker
WeasyPrint + Jinja2
Sentry
Tavily
shadcn/ui
OpenWeatherMap + Foursquare
Render + Vercel
// Engineering Judgment

What Actually Went Wrong (and how it got fixed)

Systems Thinking

Fail-closed, not fail-open
Runtime agent shielding blocks on uncertainty by default (AgentSec) — caught and reversed an earlier fail-open bug.
Judge reliability under uncertainty
Anvesh's single-LLM-judge adversarial testing evolved into AgentSec's 3-signal majority-vote verifier (dual-temp LLM + deterministic heuristic).
Graceful degradation as a design principle
Raahgir's 4 external API integrations each degrade independently to realistic defaults — the pipeline can't 500 at the data-fetch stage.
Documented bug-fix reasoning
CareFlow's multi-order truncation bug — fixed by raising token budget AND adding an explicit anti-summarization instruction, not just one or the other.

Security Practices

Fail-closed defaults
Uncertain agent actions are blocked, not allowed, by default.
Constant-time comparison
Razorpay webhook signatures verified with constant-time HMAC-SHA256 comparison.
Race-condition-safe quotas
Postgres row-locking via a single UPDATE-with-WHERE-clause instead of check-then-increment.
AST-based static scanning
Code analysis walks the AST rather than pattern-matching with regex.
.env leak, caught and fixed
A real leak across 2 repos — now a standing pre-commit habit, not just a one-off fix.