Work

Professional experience and research projects.

Atomicwork

AI Research Engineer

Atom: Universal Agentic Service Management

  • 55% reduction in response latency (45s → ~20s) and 3x throughput improvement
  • Architected offline Evaluation Pipeline & Regression Testing Suite, calibrating automated evaluators against human annotations
  • Engineered CUA capabilities into conversational automation for autonomous browser-based task execution
  • Prompt Engineering and Automated Prompt Optimisation Framework to improve agent persona consistency and tool-usage accuracy

Agentic IGA: Identity Governance & Administration

  • Orchestrator-specialist multi-agent pipeline for identity lifecycle — intent parsing, policy recommendation, JIT provisioning, autonomous anomaly triage, human-in-the-loop checkpoints

Agent Harness: Multi-Agent Runtime

  • Composable skills layer with enterprise integrations as single-purpose MCP servers
  • Deterministic YAML workflow orchestration with inter-agent state handoff and subagent spawning
  • Gateway-based harness with full context isolation — scoped memory, isolated workspaces, per-agent model routing

SuperAGI Research

AI Research Engineer

Multi-Agent Systems & AgentBuilder

  • Suite of AI agents for agentic CRM and autonomous task execution (SuperSales, SuperMeet, SuperSupport)
  • In-house Multi-Agent Orchestration Framework — hierarchical, parallel, and sequential communication for stateful interactions
  • Robust context management and agent handoff mechanisms, including voice-enabled computer-use agents within multi-agent AgentBuilder

SuperCoder 2.0: AI Pair-Programmer

  • Five-phase autonomous software development workflow with fine-tuned LLMs
  • Language-agnostic codebase indexer using LSP and vector-embedded ASTs
  • 34% on SWE-Bench Lite — Top #3 globally, #1 Open Source

AUTONODE: Cognitive GUI Automation

  • Co-led self-operating web navigation for cognitive RPA and dynamic information retrieval
  • DoRA self-training module, RAG integration, site-map pre-traversal into the Large Action Model

VEagle & V-Zen: Multimodal Models

  • Q-Former integration with architectural refinements for multimodal representation learning
  • Contextual grounding with DINO detector for efficient GUI understanding

ISRO — Space Applications Center

Summer Research Intern

Deep learning for geospatial applications. Fine-tuned LLMs and Foundation Models on curated geospatial data for natural language-driven navigation systems. Mentored by Scientist/SE Pankaj Bodani.

NTNU — Norwegian Univ. of Science & Technology

Research Intern

Attention-based transformer and Audio Spectrogram Transformer models for heart murmur detection & classification, with explainable AI techniques.

Side Projects & Research

Geometric Deep Learning

Graph neural networks and non-Euclidean representation learning for structured data.

Dense Pose Modelling via WiFi Signals

WiFi CSI-to-GAN pipeline for human pose estimation from channel state information — generating dense body-surface mappings without cameras.

SOAR: Aerial Imagery Detection

Fine-tuning State Space Models with programmable gradients for small-body object detection in aerial imagery.