● Founding AI & ML Engineer · Shinobi Security

I build multi-agent systems
for complex work.

I lead and build multi-agent harnesses, durable execution, and governed action. At Shinobi Security, I apply that work to offensive security and other high-stakes workflows in regulated environments.

MSc · MBZUAI AAAI 2024 UAE Golden Visa Dubai, UAE
View selected work → Email me
Next / operating model
Positioning

One system. Two responsibilities.

Give an agent the capability to test real attack surfaces, then constrain that capability with explicit policy, scope, and evidence.

OPERATE Offensive agents

Autonomous systems that execute security testing across complex enterprise environments.

  • Hierarchical multi-agent harnesses
  • A2A, MCP, and external tool communication
  • Durable memory, planning, and delegation

GOVERN Reliable execution

Controls that keep powerful agents inside approved boundaries and make every action inspectable.

  • Scope adherence and least-agency controls
  • Sandboxed execution and human approvals
  • Observability, evaluation, and auditability
Trajectory
Portrait of Abhishek Gehlot

Selected experience.

Research became production systems. Production systems gained the controls needed to operate safely in high-stakes environments.

01Jun 2023 — Present
Founding AI & ML Engineer
Shinobi SecurityUAE
  • Own zero-to-one technical engagements for Tier-1 fintech, banking, telecom, media-streaming, and super-app teams, from executive requirements and product workflows through deployed systems.
  • Design and build hierarchical offensive-security agent systems spanning orchestrators, supervisors, specialist teams, durable memory, MCP, and A2A communication across web, mobile, cloud, and network surfaces.
  • Delivered systems that ranked as the top-performing approach in the majority of client internal assessments against competing products and manual workflows.
  • Shifted security testing from quarterly or annual pentests to continuous testing tied to release cycles, raising vulnerability coverage to 99.9% across tested application attack surfaces.
  • Combined Neo4j attack-surface intelligence and Kubernetes deployment with scope controls and Arize Phoenix observability and evaluation for reliable, inspectable execution.
Earlier experienceMBZUAI · Dubai Police
02Aug 2021 — May 2023
Graduate Researcher
Intelligent Visual Analytics Lab, MBZUAIAbu Dhabi
  • Developed and optimized object-detection and segmentation methods for aerial and satellite imagery, contributing to published research in semi-supervised open-world detection.
  • Turned PyTorch and TensorFlow research prototypes into reusable libraries for continued experimentation and production adaptation.
03Aug 2022 — Oct 2022
Artificial Intelligence Intern
Dubai Police — Gen. Dept of AIDubai
  • Used geospatial and demographic analysis to optimize Smart Police Station coverage, improving response times by up to 35% within the targeted-zone analysis.
Research

Research, in public.

Five publications across open-world detection, self-supervised learning, and low-resource NLP, including AAAI 2024, an IEEE i-PACT Best Paper, and an ACL workshop oral presentation.

Semi-Supervised Open-World Object Detection
Mullappilly S., Gehlot A. S., Anwer R. M., Khan F. S., Cholakkal H.
AAAI 2024
Four more publicationsACL · IEEE · Springer · arXiv
An Empirical Study of Self-Supervised Learning Approaches for Object Detection with Transformers
Karthik G., Mullapilly S., Gehlot A.
arXiv
MuCoT: Multilingual Contrastive Training for Question-Answering in Low-resource Languages
Karthik G., Gehlot A., Mullapilly S.
ACL 2022 · Oral
Refinement Image Regeneration
Gehlot A., Kumar A., Sampson S. A.
IEEE i-PACT 2021 · Best Paper
Market Trend Predictions using RNNs and Sentiments
Abhayankar V., Gehlot A., Gupta P.
Springer 2020
Selected Projects

Independent builds.

Consumer products used to test how agentic systems behave outside enterprise constraints.

Project Guacamole

AI Wealth Manager

A reasoning layer that turns holdings, risk appetite, and market context into a portfolio strategy a person can inspect.

Scenario 04Balanced growth
Equity68
Fixed21
Cash11
PythonReflexGemini
Mobile App · In Progress

AI Health Coach

A cross-platform coach that converts daily health signals into specific, adaptive guidance instead of generic advice.

Weekly cadence5 / 7

Recovery load reduced after the plan adapted.

FlutterGeminiHabit modeling
Engineering Depth

How I engineer agent systems.

Three connected disciplines: coordinate capable agents, preserve reliable state, and govern every action.

Multi-Agent Harnesses

Hierarchical orchestrationAgent communication & tool interfacesPlanning & delegationMixed-model routing

Durable Intelligence

Context & memory managementDurable state & resumabilityLong-horizon failure recoveryObservability & evaluation

Governed Action

Scope adherence & least agencySandboxing & action isolationHuman approvals & interventionIdentity, authorization & auditability
GitHub Activity

Work, in motion.

Live contribution activity from my GitHub profile. Private work is represented only as anonymous daily contribution counts; repository names, code, commits, and organization details are never fetched or displayed here.

View @Abhishekq10 on GitHub
Loading live contribution data… Rolling 12 months
Connect

Always open to a thoughtful conversation about AI systems and security.

Email LinkedIn