KR

Karthik Ragulan

Dual Degree, Indian Institute of Technology Kharagpur · Class of 2027

Bangalore, India

I'm a Dual Degree student at IIT Kharagpur building systems where AI agents cooperate, check each other's work, and stay fast and reliable at real data scale.

At Glance (InMobi Group) I shipped three production systems: Apollo, a ReAct sales agent over 14M+ brands that cut query latency 30×; Arrakis, a no-code fine-tune console fitting 27B-param models on a single GPU; and DRAPE, a multi-LoRA VLM pipeline extracting 45+ catalogue attributes per garment image.

My latest personal projects are Flock — a shopping agent grounded in live catalogues across 59 verified merchants — and NightShift, an autonomous on-call agent that turns a GitHub bug label into a tested draft PR. I compete too: Gold at Deloitte OpenSoft 2025 and Silver at Inter IIT Tech Meet 14.0.

Open to SDE / AI-ML opportunities

Karthik Ragulan

↳ drag · hover to trace · click to jump

01 — By the numbers

At a glance.

30×

Latency cut on Apollo — 96s → 3s (p50) across 292 traced production runs

27B

Parameter model fit on a single 24 GB GPU via QLoRA in Arrakis

45+

Structured catalogue attributes extracted per garment image with DRAPE

59

Verified merchants grounded in Flock with zero invented claims

2

National-level medals — Gold (Deloitte OpenSoft) & Silver (Inter IIT 14.0)

93

Backend tests on Arrakis — ruff and mypy clean, zero data egress

02 — Work

Selected Work

Three production systems built at Glance (InMobi), two personal agents, and two national-level hackathon wins — both with full case studies.

SDE InternMay 2026 – Aug 2026

Apollo — Agentic Sales Analyst

Glance · InMobi Group · SDE Intern

A ReAct agent over 7 governed tools answering plain-English sales queries on 14M+ brands in Cloud Spanner. Cut median latency 96s → 3s (30×) and verified across 292 traced production runs.

  • ▸Built a ReAct agent over 7 governed tools answering plain-English sales queries on 14M+ brands in Cloud Spanner
  • ▸Cut median latency 96s → 3s (30×) by eliminating redundant model reasoning and indexing hot paths into 0.24s seeks
  • ▸Made safe by construction: read-only SELECT-only queries, 100-row caps, a 10-round loop bound, per-rep audit traces
  • ▸Verified over 292 traced production runs (p50 3s, p90 9s); added an agentic-commerce readiness score for reps
PythonReActCloud SpannerGCPMulti-AgentLangGraph
SDE InternMay 2026 – Aug 2026

Arrakis — Self-Serve Fine-Tune Console

Glance · InMobi Group · SDE Intern

A no-code console turning a raw CSV into a trained model on a single 24 GB NVIDIA L4, exposing 22 models. Fit a 27B model on one GPU with QLoRA across VLM, CNN and LLM modes — at ~$0 idle cost.

  • ▸Shipped a no-code console turning a raw CSV into a trained model on a single 24 GB NVIDIA L4, exposing 22 models
  • ▸Fit a 27B-param model on one 24 GB GPU with QLoRA (frozen 4-bit base + 1% adapter) across VLM, CNN and LLM modes
  • ▸Split CPU control plane off on-demand GPU jobs (Cloud Run min-0, auto-deleted Cloud Batch VMs) for ~$0 idle cost
FastAPIReactQLoRAUnslothTRLCloud RunCloud BatchGCP
SDE InternMay 2026 – Aug 2026

DRAPE — VLM Catalogue Enrichment

Glance · InMobi Group · SDE Intern

Fine-tuned three LoRA adapters on Qwen2.5-VL-7B to extract 45+ structured attributes from a single garment image. All adapters served off one AWQ-INT4 base via vLLM multi-LoRA with 5-prompt GPU batching.

  • ▸Fine-tuned three LoRA adapters on Qwen2.5-VL-7B to pull 45+ structured attributes from a single garment image
  • ▸Served all adapters off one AWQ-INT4 base via vLLM multi-LoRA, batching 5 prompts into one GPU forward pass
  • ▸Kept non-apparel items out of the pipeline via FunctionGemma-270M and a deterministic 160-term allowlist
Qwen2.5-VL-7BvLLMLoRAAWQ-INT4QLoRAKubernetesGCP
Personal BuildAug 2026

Flock — Multi-Merchant Shopping Agent

Personal Project

A shopping agent grounded in live UCP catalogues across 59 verified merchants — a Gemini coordinator plans, deterministic code enforces every hard constraint, with mission boards and multi-merchant bag persisted in Cloudflare D1.

  • ▸Built a shopping agent grounded in live UCP catalogues across 59 verified merchants, with no invented claims
  • ▸Split planning from enforcement: a Gemini coordinator plans, deterministic code enforces every hard constraint
  • ▸Searched merchants concurrently while streaming evidence; a 10-minute discovery cache bounds catalogue walks
TypeScriptGeminiCloudflare WorkersCloudflare D1Drizzle ORM
Personal BuildAug 2026

NightShift — Autonomous On-Call Agent

Personal Project

An event-driven agent that turns a labelled GitHub bug issue into a tested draft PR via Gemini 3 Flash. Autonomy is bounded behind a policy gate — repo allowlist, dual-label opt-in, 3-file cap, zero merge authority.

  • ▸Deployed an event-driven agent that turns a labelled GitHub bug issue into a tested draft PR via Gemini 3 Flash
  • ▸Bounded autonomy behind a policy gate: repo allowlist, dual-label opt-in, 3-file cap, and zero merge authority
  • ▸Capped planning context at 6 files / 18,000 chars and authenticated as a GitHub App with HMAC SHA-256 checks
PythonCloud RunVertex AIFirestorePub/SubGeminiGitHub App
Silver MedalNov 2025 – Dec 2025

Causal Analysis RAG System

Inter IIT Tech Meet 14.0

A multi-agent RAG system that iteratively retrieves across 3 hops until evidence is complete, reranks with a ColBERT + Cohere + Voyage ensemble, and remembers prior context via GraphRAG.

  • ▸Built a multi-agent RAG system across 5 FastAPI microservices, deployed via Docker Compose for horizontal scaling
  • ▸Architected a 3-hop LangGraph retriever with LLM sufficiency checks, iteratively refining queries until evidence is complete
  • ▸Combined ColBERT, Cohere, and Voyage into a weighted ensemble reranker, surfacing the top-10 most relevant evidences
LangGraphFastAPIColBERTCohereVoyage AIZep CloudPresidioDocker
View Case Study
Gold MedalFeb 2025 – Apr 2025

DeloConnect — AI Employee Wellness Platform

Deloitte OpenSoft 2025

A 5-agent LangGraph workflow that turns activity, leave, performance, and sentiment signals into structured HR reports — plus a counseling chatbot that knows when to escalate.

  • ▸Built a LangGraph fan-out/fan-in workflow across 5 agents (activity, leave, performance, rewards etc.) into HR reports
  • ▸Designed a counseling chatbot with an Agno/OpenAI decision agent triggering topic changes, HR escalation, or session end
  • ▸Built a Gemini summarizer converting escalated chats into structured HR reports flagging concerns, sentiment, and action items
LangGraphAgnoOpenAIGeminiFastAPIMongoDB AtlasWebSocketsIsolation Forest
View Case Study

03 — Recognition

Wins

Competing on national stages, and bringing back the medals to show for it.

🥇

Gold Medal

Deloitte OpenSoft 2025

🥈

Silver Medal

Inter IIT Tech Meet 14.0

04 — Toolkit

Skills & Stack

The languages, frameworks, and infrastructure I reach for most.

Languages

Python · TypeScript · JavaScript · SQL · Bash

AI / ML

PyTorch · HuggingFace Transformers · Unsloth · TRL · vLLM · LoRA & QLoRA Fine-Tuning · LangGraph · RAG · ReAct & Multi-Agent Systems · Vertex AI (Gemini) · OpenAI API

Frameworks

FastAPI · Next.js · React · React Native (Expo) · Flask · Drizzle ORM

Data Stores

PostgreSQL · MongoDB · Cloud Spanner · Firestore · Cloudflare D1 · Qdrant · ChromaDB · FAISS · Neo4j

Cloud & DevOps

GCP (Cloud Run · Cloud Batch · Pub/Sub · Vertex AI · GKE · Secret Manager) · Cloudflare Workers · Docker · Kubernetes · GitHub Actions · Modal · Nginx

Education

2027

Dual Degree (B.Tech + M.Tech)

Indian Institute of Technology Kharagpur · CGPA 7.35

2022

AISSCE (Class XII)

Narayana Junior College, Bangalore · 91.0%

2020

AISSE (Class X)

Amara Jyothi Public School, Bangalore · 89.4%

05 — Contact

Let's build something.

Open to SDE and AI/ML internships — or just want to talk about multi-agent systems? Reach out.