Senior Software Engineer
Adya.ai
Autonomous Multi-Agent Platform (SAI)
- Architected the supervisor core of SAI, an autonomous multi-agent delivery system that turns a plain-language request into delivered artifacts — classifying intent, generating multi-phase plans, delegating to specialist agents, and running independent steps in parallel through a DAG engine with dependency resolution, retries and cascade-failure handling.
- Re-architected the skill system into a central skill registry — metadata indexed in a vector database, raw definitions in MongoDB — equipping 1,400+ skills on demand at runtime and auto-installing new capabilities from an online marketplace when no local match exists.
- Engineered a 3-layer context-management system (Auto Compact) — cross-request, per-call and per-tool-call compaction — so long agent-building sessions run unbounded without hitting LLM token limits.
- Built HAA (Human Assistant Agent) on a per-workspace VM context layer that stores and retrieves workflow context, memory and session summaries — letting the assistant auto-answer user questions on their behalf.
- Added human-in-the-loop control (plan approve / modify / reject, mid-run questions) and multi-channel engagement — WhatsApp, Email, In-App and Web Push — so users can answer, check status and steer long runs from anywhere.
AI Workflow Automation (AIA)
- Re-architected the natural-language-to-workflow builder into a master-LLM recursive tool loop that orchestrates specialist sub-agents (requirement gathering, component prediction, flow connection, validation & repair), grounded in a 750+ component catalog (including MCPs) via vector retrieval.
- Bootstrapped the knowledge base by running real use-cases through AIA's own synthesis engine to generate ~2,000 domain-segmented reference workflows, fed back as grounding data to strengthen the component ontology, retrieval and cold-start learning.
- Streamed every node and edge live to the canvas on create and update paths, with structural validation and one-shot automated repair of malformed flows.
Self-Improving Learning Layer (ESLL)
- The platform learns from its own results and gets better with every build. Architected ESLL, a self-improving, event-sourced learning layer for generative AI: it observes each decision and its context, extracts durable knowledge from behavioural telemetry, injects that knowledge into future decisions under a token budget, and attributes outcomes back to the knowledge that shaped them — promoting what works, demoting what doesn't, with no gradient training.
- Deployed ESLL on the AIA builder to improve workflow synthesis — using materialized projections for credit assignment and demand-signal extraction to surface capability gaps as a product-roadmap input.
Backend, Data & Platform Engineering
- Designed a no-code ETL platform across major databases, cloud storage and SaaS sources, then evolved it into a fully agentic data pipeline driven by natural-language conversation.
- Built real-time collaboration and streaming with WebSockets and Server-Sent Events for live concurrent editing and low-latency updates.
- Cut ~30% RAM under peak load through targeted indexing, Redis caching and MongoDB connection-pool tuning; handled burst traffic with Go routines and channels at stable latency.
- Integrated multiple MCP servers and extensible components for modular, configurable agent workflows at scale.
Cloud, Infrastructure & Leadership
- Architected Cloud Studio, a unified PaaS / SaaS deployment platform for AI agents; managed multi-cloud infrastructure hands-on — VMs and AKS on Azure, Compute VMs on GCP, Bedrock LLM access on AWS — with Azure AI Foundry for model and key management.
- Maintained Dockerised service pipelines, Jenkins deployments and Nginx configuration for consistent environments.
- Mentored engineers and drove architectural decisions; represented Adya.ai at launch events and workshops, and served as Lead Jury at the SECE hackathon.