Kailash Kumar Resume

AI & BACKEND ENGINEER · BANGALORE → OPEN TO RELOCATE

I build AI agents
that do real work.

Tech lead of Adzump-AI at Modlix, where a team of three AI agents sets up Meta & Google Ads campaigns end to end. Built on the auth, scheduling, and integration backend I spent the previous three years building.

See how it works ↓ Email me

GitHub ↗ Resume (PDF) LinkedIn ↗ X ↗

01 - THE WORK · ADZUMP-AI · MODLIX · OCT 2025–PRESENT

What I've built,
in plain terms.

Adzump automates ad campaigns for Meta & Google. I lead its AI core. Employer code is private, so the systems are described here, and a clean-room open-source build is in progress below.

1

Three agents,
one campaign.

Give Adzump a business, and a team of agents gets to work: a Product Analyst researches competitors on the live web, a Vision Analyst reviews logos and creatives, and a Profile Writer turns what they found into a product profile.

I designed the framework that coordinates them: a router that treats whole agents as tools, decides who works, and runs them in parallel when it can.

custom multi-agent framework · 9 tool groups · Claude, OpenAI o-series & DeepSeek · prompt caching · context management

ONE ORCHESTRATOR RUN
  1. Decides which agents work on the request router
  2. Researches competitors on the live web product analyst
  3. Reviews logos & creatives vision analyst
  4. Writes the product profile profile writer

2

A conversation that ends
with a live campaign.

ChatV2 walks an advertiser from "I want to run ads" to a configured campaign: it collects the details, confirms locations, predicts a budget, picks the account, and asks for a final yes, streaming every token as it thinks.

LangGraph + LangChain · 9-node state machine · SSE streaming · gpt-4o-mini + Gemini

CHATV2 · SETUP FLOW
  • collect data
  • confirm location
  • predict budget
  • select account
  • confirm ✓

3

Models that keep
the agents honest.

Agents touching real ad budgets can't run unsupervised. Every session is scored on tool success, user satisfaction, retries, and token cost, and failures are mined for patterns. A LightGBM model forecasts ad performance from keyword embeddings, and a feedback loop remembers every keyword a user rejected so it's never suggested again.

session-scoring evals · failure-mode extraction · LightGBM forecasting · linear-regression budget · RAG on pgvector

EVERY AGENT RUN IS SCORED ON
  • tool success
  • satisfaction
  • retry & undo penalties
  • token efficiency

SEPARATE RULE-BASED METRIC EVALUATORS

  • CTR
  • CPL
  • conversion rate
  • quality score

4

The unglamorous backend
that makes it possible.

Before the agents, I spent three years building the platform they now run on: a Java/Spring Boot backend with an HA-clustered scheduler that never double-fires, auth used by multiple client apps (RBAC, SSO, OAuth 2.0), and an integration layer of 250+ API endpoints across Meta, Google, WhatsApp, and Zoho, with idempotency and retry/backoff, because third-party APIs fail.

Java · Spring Boot · Quartz + JDBC JobStore · Redis · MySQL/PostgreSQL · idempotency + retry/backoff

250+API endpoints integrated
3+years in production
0double-fired scheduled jobs by design
FOR THE ENGINEERS The architecture, one level deeper
Multi-agent orchestrator: core loop routing to three specialized sub-agents ORCHESTRATOR core tool-use loop · parallel tools AGENT ROUTER PRODUCT ANALYST competitor discovery VISION ANALYST creative selection PROFILE WRITER context → summary PROVIDERS: CLAUDE · OPENAI o-SERIES · DEEPSEEK
click a node to inspect it

02 - EXPERIENCE

One company,
two chapters.

Modlix · NoCode SaaS

BENGALURU · NOV 2022 – PRESENT

Tech Lead · Adzump-AI

OCT 2025 – PRESENT

Everything in §01: the multi-agent framework, the LangGraph engine, the forecasting models, and the eval loop that grades them.

▲ PROMOTED TO TECH LEAD

Backend Engineer

NOV 2022 – SEP 2025
  • HA-clustered job scheduler (Quartz + JDBC JobStore, DB-lock coordination) for the platform's Java/Spring Boot microservices, with no double execution.
  • Authentication & authorization (RBAC, SSO, OAuth 2.0 social login) across multiple client applications.
  • The integration layer: 250+ endpoints across Meta Ads, Google Ads, WhatsApp, Zoho, with idempotency and retry/backoff.
  • Redis-backed token & caching layer, and a lead pipeline that auto-creates adsets/ads from Google campaign data.

03 - OPEN SOURCE & PROJECTS

Beyond Adzump.

multi-agent-orchestrator

IN PROGRESS

A clean-room, employer-IP-free build of the pattern from §01: core loop, agent router, two sub-agents, and an eval harness that benchmarks why a routed multi-agent system beats a flat tool list. Repo and live demo land here.

github.com/KailashKumar28/multi-agent-orchestrator

Smart-Asset · real-estate CRM

2022–2024

50+ REST APIs (Spring Boot, MySQL) and the Inventory, CostSheet, and Booking modules, streamlining sales workflows and reducing manual errors.

04 - SKILLS & ABOUT

The shortlist.

Mechanical engineer by degree (IIT Gandhinagar, 2018–2022), backend engineer by trade, and since late 2025 the tech lead putting agents in front of real ad budgets. The through-line: systems that keep working when nobody is watching.

Based in Bangalore, open to relocation (visa sponsorship required). Looking for AI/ML Engineer, Founding Engineer, Forward-Deployed, or Senior Backend roles at AI startups.

Languages
Java · Golang · Python · TypeScript · JavaScript · C++
AI & LLM
Multi-agent systems · LangGraph/LangChain · Claude, OpenAI, Gemini, DeepSeek · RAG (pgvector) · prompt caching · SSE
Backend
Spring Boot · Spring Security · WebFlux · FastAPI · JPA · jOOQ · RabbitMQ · microservices
ML & Data
LightGBM · scikit-learn · sentence-transformers · pandas · NumPy
Databases
PostgreSQL · MySQL · MongoDB · Redis · ClickHouse · InfluxDB
Cloud & Ops
AWS · OCI · Docker · CI/CD · Grafana · observability · distributed systems

Building agents on real workloads?
Let's talk.

Email me LinkedIn

BANGALORE, IN · OPEN TO RELOCATION (VISA SPONSORSHIP REQUIRED)