Sunil Gudivada.

NIYO

A local-first autonomous office where specialized Strategy, Builder, Researcher, Reviewer, and Memory agents collaborate on shared missions while deterministic host controls keep ownership, actions, evidence, revisions, and approvals traceable.

• Plan: A human defines goals, constraints, permissions, and acceptance criteria; a planning agent decomposes the mission and assigns named owners.
• Propose: A Node.js host supplies bounded context through a JSONL bridge to Python TinyWorld and TinyPerson agents, with GitHub Copilot SDK generating role-driven action proposals.
• Govern: The host validates authorization, schemas, budgets, revisions, mission state, and approvals before controlled Browser, Documents, Work IQ, or Worktree adapters can act.
• Verify: Current evidence, independent review, verification, and human acceptance gate delivery; SQLite stores mission state, provenance, approvals, and execution history.

Sutradhar

An evidence-grounded advisor-to-project discovery platform that replaces keyword-only matching with semantic retrieval, LLM reasoning, explainable recommendations, and portfolio-level balancing. Reached 1,700+ advisors and placed in the Top 10 projects for Science Fair Shift 2 at Microsoft Hyderabad Hackathon 2026.

• Collect: Refresh advisor expertise, interests, and Innovation Studio project metadata without requiring a dedicated backend API.
• Retrieve and reason: Use embeddings and semantic search to identify candidates, then apply evidence-grounded LLM evaluation to review fit.
• Balance and explain: Apply capacity-constrained optimization without forcing weak assignments and generate inspectable "why you fit" rationales.
• Deliver: Precompute recommendation artifacts into a VibeHub static experience with no online model calls during browsing.

Micro Agents for Learning

A reusable architectural pattern that decomposes a monolithic learning agent into focused capabilities, enabling Coach, Teach, and Deep Learning experiences while improving scalability and team autonomy.

SEVAL Quality Ecosystem

Validator, analyzer, and regression-classifier tools that moved agent quality earlier in the lifecycle and enabled data-driven release decisions across more than 700 evaluation queries.

Loki x Viva Learning

Secure, traceable platform foundations used by more than five major features. The integration moved partner onboarding from weeks to days and reduced debugging time by approximately 60%.

Copilot Academy & Agent Prototypes

General-availability learning experiences alongside explorations in orchestration, memory, deep research, adaptive learning cards, and MCP. Hackathon work produced nine ideas, five prototypes, and two Pinnacle selections.

Agent Pulse

An AI-agent observability and release-confidence platform that retains evaluation evidence and historical context so teams can diagnose regressions, understand behavior changes, and make informed release decisions.

• Ingest: Aggregate Civil and SEVAL outputs plus uploaded development conversations into request-, feature-, and segment-level signals.
• Analyze: Persist run history for longitudinal analysis and apply anomaly detection, surfacing 188 anomalies during project analysis.
• Decide: Compare releases across quality, reasoning, latency, cost, prompts, workflows, and feature changes through searchable dashboards.
• Operationalize: Run as a diagnostic layer over Civil and SEVAL with an Azure DevOps repository and CI pipeline.

Teams Meeting Digital Twin

A GraphRAG-powered digital twin that models meeting participants as differentiated personas and simulates realistic interactions to study behavior, decision patterns, and human-agent collaboration before customer-facing validation.

• Ground: Structure participant and meeting context through a GraphRAG knowledge layer.
• Model: Convert grounded context into representative personas with distinct goals, behaviors, and decision perspectives.
• Simulate: Use TinyTroupe to run multi-agent meeting interactions and analyze resulting conversations and outcomes.

Persona Research & Customer Delight

A repeatable research and assessment workflow for evaluating AI experiences against realistic customer goals, behaviors, and expectations instead of relying only on generic test prompts.

• Define: Create representative personas and journey scenarios with distinct goals, constraints, behaviors, and expectations.
• Exercise: Run journeys from each persona's viewpoint through a repeatable agent-based evaluation process.
• Assess: Identify strengths, failure patterns, and customer-impact gaps and translate them into prioritized product improvements.
• Communicate: Produce persona-research artifacts and walkthroughs for engineering and product review.

System Design Studio

A visual learning studio for exploring system design concepts and architecture solutions. The product helps engineers reason about scalability, distributed systems, and design trade-offs through a focused, interactive experience.

Conceived before the mainstream GPT era; later evolved with AI-assisted curation to organize and enrich its system-design problem library.

Robo Rumble

Led core agent architecture work among 300+ global participants for a system in which a cloud orchestrator coordinates multiple Raspberry Pi robots to solve shared tasks through specialized hardware capabilities and continuous feedback.

• Delegate: Extract and segment a puzzle or task at the orchestrator, then assign subtasks to specialized robot agents.
• Execute: Use Semantic Kernel plugins to expose hardware capabilities while Raspberry Pi clients use cameras, sensors, navigation, and OpenCV.
• Coordinate: Exchange tasks and results over WebSockets between the cloud orchestrator and field robots.
• Replan: Feed outcomes back to the orchestrator for evaluation, scoring, and iterative task planning.

VoteMate

Supported an NGO-focused, Microsoft-incubated non-partisan voting companion that helps users understand ballots, candidates, and public issues while emphasizing accessible information, privacy, and trust.

• Personalize: Capture location and voting preferences through React to initiate a personalized ballot and candidate-information experience.
• Retrieve and reason: Connect Python services to Azure AI Foundry and GPT capabilities; use RAG-oriented agents to combine voting knowledge, candidate data, and public-web research.
• Present: Deliver clear, contextual ballot and candidate information through the web experience.
• Contribute: Support architecture, AI direction, engineering reviews, front-end planning, and custom agent-engine work.