DATAOPTIMIZATIONReads catalogues, spec sheets, drawings and any-shape Excel into a structured product master, with the source of every value.
Selva RagulKumarBalakrishnan
I build systems where AI agents do the work and people review the exceptions.
Lead Engineer (CTO Candidate) on ZAG at Pioneerwork, Tokyo.
Work
Product
Lead Engineer
AI Engineer
Research Intern
Researcher
Programmes
2024 cohort
GSAP AI CTO 2026
Study and awards
B.E. CSE
BS Data Science
Young Scientist India
My work.
ZAG makes product information readable by people and AI agents, keeps it that way, and collects it from whoever owns it.
DATAOPTIMIZATIONAGENTICOPTIMIZATIONMeasures how AI assistants describe a brand's products, finds what they get wrong or leave out, and fixes the sources.
FOR EXCELAn Excel add-in that fills retailers' product registration templates from catalogues and spec sheets.
98.4%
Accuracy from the agents alone. Measured across about 120 production catalogue jobs, before any person touches the data.
100%
Visit withzag.com ↗Accuracy delivered. An AI-assisted annotation process I built finds the remaining errors fast, and every one is fixed before the data ships.
Autonomous zeroth responder
A drone that becomes the network when the towers go down.
It flies in, becomes the Wi-Fi, scores every SOS with on-board AI and relays it over LoRa to the ground.
NGN Lab, Anna University, 2023–2025. Published patent application, inventor.
Watch the drone fly
Research.
Three IEEE papers. Two came straight out of the drone.
-
IEEE Internet of Things Journal, 2026
IEEE Xplore ↗RescueNav: AI-Powered AAV Trajectory Planning With LoRa Connectivity for IoT-Based Disaster Communications
4× signal strength, 42% more throughput in real flights.
Raja, Bhooma, Kumar, Dev. Vol. 13, Issue 9
-
IEEE ICC 2026, Glasgow
IEEE Xplore ↗FLAIR: Fuzzy Logic-Based Aerial Intelligent Routing for Efficient Disaster Relief Operations
92.7% LoRa line of sight with over 50% less CPU.
Raja, Kumar, Bhooma, Essaky, Veeramani, Dev
-
IEEE Transactions on Consumer Electronics, 2026
IEEE Xplore ↗Decentralizing Land Registry Infrastructure: Consumer-Run Shared Blockchain for Sustainable Agricultural Applications
Sharded blockchain on consumer nodes; 95% land-value accuracy.
Raja, Senthivel, Kumar, Kundu, Quasim, Dev
Proof.
Everything on this page, checkable.
Indian patent application, published
A System for Autonomous UAV-Assisted Disaster Response and Method Thereof
- Application
- 202541032226 A
- Filed
- 1 Apr 2025
- Published
- 2 May 2025, Journal No. 18/2025
- Claims
- 8
- Applicant
- Anna University, MIT
Inventors: Dr. R. Gunasekaran, Selva Ragul Kumar, J B Mugundh, Theodore Philip, Hirthik Raj A, Selvam Essaky.
Patent Office Journal (PDF) ↗
Where I've done the work.
Four years building through college, then straight into production AI.
Apr 2026 – now
Lead Engineer (CTO Candidate)
Pioneerwork Inc. (Techstars 2024), Tokyo
- Lead engineering across all three ZAG products, on Azure.
- 98.4% → 100% accuracy across ~120 production jobs.
- Upstream data work on SI projects worth tens of millions of yen.
Jun 2025 – Apr 2026
AI Engineer
Thirdwave Corp., Tokyo
- Tech lead on Aviary: local models and a ChatGPT-style chat.
- Fine-tuning proof of concept on CUDA.
Jun 2023 – Jun 2025
Researcher
NGN Lab, Anna University, Chennai
- Led ResQ-Link from design to flying prototype.
- LLM and RAG research on vehicle platooning.
Jun 2024 – Jul 2024
Research Intern
Zoho Corp., India
- Graph RAG for Site24x7 support bots.
Education and awards
- B.E. Computer ScienceAnna University, MIT2021–25
- BS Programming and Data ScienceIIT Madras2021–26
- GSAP, AI CTO courseERA Global and JETRO2026
- Young Scientist India, top 10Space Kidz India2019
Toolkit
Agents
- Claude Code
- Multi-agent workflows
- Temporal
- LLM and RAG
- Azure AI Foundry
Backend
- Python
- FastAPI
- PostgreSQL
- Pydantic
- Neo4j
- Pandas
Cloud and ML
- Azure
- AWS
- Google Cloud
- PyTorch
- TensorFlow
- CUDA
Hardware
- Pixhawk
- Raspberry Pi
- LoRa
- OpenWRT
- MAVLink
Also
- C
- C++
- Java
- JavaScript
- React
- Node
- Angular
How I build.
One engineer, a swarm of agents. I run it top-down: I set the goals, agents do the work, and nothing ships without proof.
- 01You set the goal.
Each morning I set what must, could and might get done. That is the only thing I write by hand.
- 02An orchestrator plans it.
It breaks the goals into one-session tasks, writes a markdown brief for each, and pulls context from my knowledge graph of mail, meeting notes and calendar.
- 03Project agents take over.
Every project has its own agents with their own persona and rules.
- 04A swarm builds in parallel.
Dozens of Claude Code sessions implement and test at once. A watchdog restarts anything that stalls for 30 minutes.
- 05Reviewers try to break it.
A project-manager agent checks every result against the diff; adversarial reviewers send the failures back to be fixed.
- 06Only what holds ships.
Tests, staging and QA loops first. Production needs my approval, always.
1agents running
0tasks in flight