Founding Member of Technical Staff at RISA (2024-07 – Present)
Built RISA’s entire AI intelligence layer from 0→1. Sole architect and implementer of foundational agentic workflow automation.
- Designed and implemented BOSS Console - multi-agent communication architecture using the Agent-to-Agent (A2A) protocol with asyncio’s async message-passing; built an orchestrator-specialist delegation model (Router Agent → Specialists like PDF Parser / FHIR Adapter), enabling parallel agentic pipelines with structured handoffs achieving >30% zero-touch (no HITL) prior-authorisation workflows and 10× order volume scaling without linear headcount growth.
- Pioneered Model Context Protocol (MCP) as the agent harness layer - built MCP server with JSON-RPC handshake for dynamic tool discovery at agent spin-up time, eliminating context-window bloat (previously 3× token overhead from static prompt stuffing); implemented permission-scoped tool delivery so each agent instance only receives schemas relevant to its task.
- Self-initiated building LLM-Powered RPA Workflow Recorder - engineered ‘Magic Selectors’ prompt strategy: instead of feeding full DOM to LLM (context blowout), overlaid precisely cropped HTML snippets at click points for near-100% xpath selector accuracy; deployed NoVNC remote-browser infrastructure enabling non-engineers to record RPA configs, offloading 10-12 workflow configs/month from engineering and constituting a novel agent affordance for human-AI collaborative workflow creation.
- Architected vendor-agnostic AI Agent Factory (Abstract Factory + Registry pattern) supporting runtime LLM switching between GPT, Claude, Gemini, and TogetherAI; designed Adapter pattern to normalise varying I/O schemas across providers, enabling hot-swap failover and $420K/year cost avoidance by migrating from in-house LLM hosting to API-based architecture (~$7/million tokens).
- Designed quantitative evaluation framework for agentic pipeline performance: benchmarked GPT, Claude, Gemini, and TogetherAI across latency, zero-touch rate (accuracy), and cost/million tokens; established MTTR tracking via context-aware error capture (automated screenshot + DOM dump at failure point), reducing debug cycles by ~60%.
- Designed Human-in-the-Loop (HITL) Standard Operating Procedures and audit-trail infrastructure for all automated AI decisions, including permission-scoped MCP tool, masked BigQuery logs, customer-facing interpretability dashboards, ensuring human oversight and auditability at scale while creating reliable feedback.
- Reduced order ingestion from 75 → 7 minutes (10× improvement); saved 500–1,000+ hours/year in fulfillment capacity; stabilised 9-microservice platform via HPAs, zombie-process RCA, and cleanup schedules.
Software Engineering Intern at Sprinklr (2024-01 – 2024-06)
- Implemented resiliency checks for CPU under/over-utilisation in Kubernetes deployments with Teams webhook alerts; identified issues in 60%+ of deployments, impacting 40% in 95th-percentile replica-count clusters.
- Developed proactive jobs to detect Elasticsearch shard mapping discrepancies and fixed missing-replica bugs; introduced co-ordinating-node logging for search requests across ES clusters.
- Analysed and fixed 15+ cloud infrastructure issues via jstack, heap dumps (Eclipse MAT), Grafana, and KEDA HPA analysis.
Research Intern at Indian Institute of Science (2022-03 – 2023-12)
- Designed novel architecture combining Vision Transformers (ViT) and Masked Auto Encoders (MAE) for micro-expression emotion recognition; curated multi-angle facial dataset with varying emotion intensity - outperforming SOTA by +11% on SAMM and CASME2 benchmarks. This systematic architecture design and evaluation methodology directly informs my approach to building and benchmarking agentic systems.
- Conducted comparative analysis of CNN-based architectures (ResNet, VGG-16) vs Vision Transformers for classification accuracy.
Research and Development Intern at Soroco (2022-08 – 2023-04)
- Devised a rule-engine to summarise 4-10 action events into single representative event (<25% data size), generating synthetic training data for user-behaviour analysis using heuristics and GPT-3 for natural language action sequence summaries.
- Anonymised sensitive information by mapping words to domain categories, creating publicly available datasets - an early application of data-curation techniques for LLM training pipelines.
Software Engineering Intern at Sprinklr (2023-05 – 2023-07)
- Deployed storage-analysis APIs to identify unused ElasticSearch fields, reducing storage costs ~30%; built Azure Cost Analyzer with 12 drill-down views for anomaly detection and cost forecasting.
Research Intern at CEERI (2022-05 – 2022-07)
- Implemented Eulerian Video Magnification for remote plethysmography (rPPG): spatio-temporal decomposition, 0.8–1.0 Hz bandpass, 20× colour/motion magnification on UBFC-RPPG and CEERI proprietary datasets.
