Nishanth Gopinath (Shan)

I build multi-agent systems and make them fast. Now asking how federated learning survives an aggregator you can't trust.

Dublin, Ireland — open to AI/ML engineering roles and PhD positions.

orchestrator KittyKat 30s → 290ms SPEAR 5h → 15min Kastro red / blue audit ApplyAgent runs itself
Watch it fan out. Or click any node to jump to that project.

orchestrator idle · 4 agents registered · awaiting request

Work

Four systems, each one a set of agents that hand work to each other. The numbers are the part I care about.

290ms Vector search, down from 30 seconds

KittyKat

AI/ML Software Developer, Aug–Oct 2025

Built a multi-agent supervisor platform on LangGraph for a fashion-tech studio, with results streaming to a Next.js front end as they resolved. The search layer was the bottleneck: moving it onto MongoDB Atlas HNSW indexes took lookups from thirty seconds to under three hundred milliseconds. Separately traced a cold-start problem on Cloud Run to synchronous invocation and fixed it by moving the trigger path to Pub/Sub.

Also shipped the n8n automation side — a Gemini Vision pipeline that pulls structured product data out of supplier images, and moodboard autofill running on the same vector index.

Sequential scan
0 ms
HNSW index
0 ms

Same query, same corpus. Playback compressed roughly seven times.

  • LangGraph
  • MongoDB Atlas HNSW
  • Cloud Run
  • Pub/Sub
  • Next.js
  • Gemini Vision
  • n8n
15min Procurement review, down from 5 hours · 1,000+ documents a day

SPEAR

Four-agent retrieval system for procurement analysis

Tender documents arrive in the hundreds and get read by people who bill by the hour. SPEAR splits that reading across four specialised agents coordinated in LangGraph — retrieval over a FAISS index, extraction, cross-checking, and synthesis — served through FastAPI with Gemini doing the reasoning. A five-hour manual pass now takes fifteen minutes, and it holds up at over a thousand documents a day.

  • LangGraph
  • FAISS
  • FastAPI
  • Gemini
  • RAG
4 agents Orchestrator, Red, Blue, Legislator

Kastro

Hiring bias auditor · Tech Ireland National AI Challenge 2026

An auditor for automated hiring systems, built with a team and showcased at Dogpatch Labs. I own the agent layer, written on Google ADK: a Red agent that probes a screening model for disparate treatment, a Blue agent that argues the other side, a Legislator agent that grounds whatever they find in EU AI Act obligations for high-risk hiring systems, and an Orchestrator that keeps the exchange converging on evidence instead of opinion.

  • Google ADK
  • Multi-agent
  • Red teaming
  • EU AI Act
  • Fairness metrics
0 clicks Driven end to end from a Telegram thread

ApplyAgent

Job application automation, controlled from a chat window

A LangGraph state machine that drives my career-ops CLI headlessly and takes its instructions over Telegram. Paste a posting into a chat, and it parses the role, tailors the CV against it, runs the ATS check, and reports back. Built for myself, which meant it had to survive real use rather than a demo.

  • LangGraph
  • Telegram Bot API
  • Python
  • Headless CLI

Also: a health-insurance assistant built for Kota’s challenge at the Give(a)Go hackathon — FAISS retrieval over policy documents, sentence-transformers embeddings, FastAPI, containerised. And the tooling underneath my own job search: a LaTeX CV pipeline, a resume-tailoring skill, and a mandatory ATS check that refuses to let a broken PDF out the door.

Research

Privacy-preserving machine learning, and the trust boundaries people assume are covered but aren’t.

Quantum key distribution secures the channel against an eavesdropper. It does nothing about the server doing the aggregating.

QKD QKD QKD QKD hospital A hospital B hospital C hospital D aggregator sees every update
Four hospitals train locally and hand their model updates to one server. Turn the channel encryption on and watch what does not change. Every link is now secure against an eavesdropper. The aggregator still reads every update in full. Protecting the channel was never about the party sitting at the end of it, and that is the part I want to work on.

Personalised Secure Slimmable Quantum Federated Learning for healthcare

MSc dissertation · Trinity College Dublin · 73%

Federated learning for clinical data has to work across hospitals with unequal compute and unequal patient distributions, which is where slimmable networks and per-client personalisation come in. The security layer uses QKD to protect the links — and that is exactly where the interesting gap opens, because a protected channel still terminates at an aggregator with full sight of every update. That gap is what I want to spend a PhD on.

Best Paper Presenter Award

Fetal health prediction using convolutional neural networks

Published during BTech, Artificial Intelligence and Machine Learning

Classifying cardiotocography signals into fetal health categories with a CNN, on a dataset where the classes that matter clinically are the rarest ones. First real lesson in how far accuracy can be from usefulness.

Background

2025
AI/ML Software Developer, KittyKat Multi-agent platform, vector search, cloud reliability, automation pipelines.
2023
Plumb5 Integrated the ChatGPT API into analytics reporting — production LLM work before it was a job title.
MSc
Computer Science (Data Science), Trinity College Dublin Dissertation on secure federated learning for healthcare.
BTech
Artificial Intelligence and Machine Learning Published research on CNN-based fetal health prediction.

Toolkit

Agents
LangGraph   Google ADK   n8n   tool orchestration
Models
Gemini   OpenAI API   PyTorch   sentence-transformers
Retrieval
FAISS   MongoDB Atlas vector search   RAG pipelines
Services
FastAPI   Next.js   Docker   Cloud Run   Pub/Sub
Research
Federated learning   differential privacy   CNNs

Contact

If you are hiring for agent work, or supervising research on privacy in distributed learning, I would like to hear from you.

Dublin, Ireland