Human-AI Capability · Sanshar Swarm · Multimodal Agents

Harihar Thapa

I build practical AI systems that connect models, interfaces, infrastructure, human workflows, and proof loops. My current work is Sanshar: an event-driven swarm testbed for Discord, voice, vision, cloud, and local machines.

Resume PDF Capability Case Study Public Sanshar Repo Project Flows LinkedIn
H-1B AWS Support Engineer Plano, TX

What I Bring

I combine cloud operations, network debugging, teaching, frontend prototypes, and hands-on AI systems work. My work is strongest where systems need to be useful in the real world: observable, rate-limited, reversible, and grounded in source evidence.

Capability
AI workflows designed around real human work, learning, and decision support.
AI Systems
Agent orchestration, event gateways, source packets, evaluation, verifier loops.
Product Sense
Rapid prototypes shaped around a real user, not a demo script.

Selected Work

Public artifact Sanitized Sanshar repo with schemas and examples Capability case study How Sanshar maps to human-AI product loops Proof discipline Source packets, read-back, postproof, retickets Secure deploy TLS, rate limits, DNS, AWS relay operations

Sanshar Swarm

A live multi-agent systems testbed where specialized peers coordinate through event streams, local machines, Discord surfaces, and proof artifacts. The work focuses on making agents useful under real operating constraints: rate limits, memory pressure, stale context, private surfaces, and human correction.

  • Manager-Agent-Verifier packets with expected metrics, observed metrics, source refs, and postproof.
  • Event-driven Discord Gateway capture with cursor state, read-back, and reaction proof semantics.
  • Public sanitized repo showing schemas, dynamic decision records, and safety boundaries.

View flow · GitHub

Dynamic Attention and Zoom Policy

A runtime policy layer for deciding when an AI system should observe, probe, summarize, ask, act, escalate, or hold. The policy replaces static behavior with measured runtime choices across language, surface, modality, risk, privacy, model, verifier, and autonomy.

  • Expected-vs-observed metrics before and after action.
  • Reason codes for pass, partial, blocked, timeout, false positive, and false negative.
  • Replay-oriented handling for missed expectations, stale context, and noisy surface signals.

View flow

Medha Labs Chess and Vision AI

Product prototypes for model-assisted chess reasoning, board-state validation, visual context, and tutoring workflows. This work shaped the Sanshar pattern of combining model suggestions with deterministic validators, replay, and confidence scoring.

  • LLM coach and judge loops wrapped around valid game state.
  • Vision and attachment handling treated as first-class source packets.
  • Customer-facing learning flows, not only backend experiments.

View flow

Experience

AWS Support Engineer, Networking and Security

VPC, Route 53, WAF, Shield, and Network Firewall support across customer troubleshooting, packet-level reasoning, and production constraints.

Network Administrator, Montrose Hospital

Hands-on hospital network operations in Colorado, balancing reliability, security, and user-facing urgency.

Teaching Assistant and Builder

University of New Haven teaching support, applied networking, and self-directed AI systems development.

Contact

Open to human-AI capability, design engineering, applied AI systems, and agentic tooling roles.

bharihar400@gmail.com hariharsecure@gmail.com +1 (203) 901-2971 LinkedIn