Problem
Long-running assistants drift, miss surface delivery, over-trust summaries, and lose state across sessions.
Back to portfolio
Sanshar Prototype Details
A prototype that connects models to real surfaces
How my AI systems observe, decide, act, prove, and learn under real operating constraints.
Sanshar Swarm
Event-driven multi-agent operating lab
Sanshar is my prototype lab for turning frontier models into an operating system of peers: worker, reviewer, witness, verifier, and assistant-style peer. The hard problem is not "can a model answer?" It is whether a system can observe the right event, choose the right tool, act safely, leave proof, and improve without silently learning the wrong thing.
Built
Manager-Agent-Verifier packet loops, source packets, reason codes, handoffs, postproof, and peer-specific lanes.
Proof
Expected-vs-observed metrics, read-back contracts, cursor state, local canaries, and retickets for repeated failures.
Why It Matters
This maps directly to enterprise Claude adoption: tool use, evals, safety gates, workflow reliability, and customer trust.
- ObserveEvent surfaces, terminal state, files, voice, images, and peer outputs.
- PacketizeSource refs, hashes, claim type, freshness, confidence, and privacy state.
- DecideDynamic routing across risk, surface, model, verifier, and autonomy dimensions.
- ActLocal reversible work first; external or high-risk changes require approval.
- ClosePostproof, read-back, reason codes, retickets, and learning candidates.
Attention and Zoom
Dynamic policy instead of static behavior
The attention layer decides when an AI peer should zoom into a signal, hold, ask, probe, summarize, escalate, or ignore noise. It turns "autonomy" into a measurable decision: what changed, why it matters, what context is fresh, what action is allowed, and what proof will close the loop.
Dynamic Dimensions
Language, modality, surface, user intent, urgency, privacy, trust, resource pressure, model choice, and autonomy level.
Failure Handling
False positives, false negatives, stale context, missed promises, timeouts, and repeated-output misses become reason-coded evidence.
Resource Gates
Memory pressure, rate limits, active agents, latency budget, and tool availability shape whether the system acts or backs off.
Learning Gate
Runtime hypotheses stay separate from durable memory, canon, knowledge graph, and training candidates until verified.
Why It Matters
Enterprises need agents that know when to act, when to ask, and how to explain the decision afterward.
- SignalDetect what changed and classify the event.
- ZoomSelect the smallest fresh context window.
- GateCheck risk, permission, resources, and privacy.
- ProofCompare expected vs observed and close or reticket.
Discord Surface Layer
Human-facing event surface with proof
Discord is treated as an operational interface, not a loose chat stream. Messages, attachments, reactions, posts, and read-back each have separate proof semantics. This lets peers coordinate without confusing "saw a message" with "completed the promise."
Event Capture
Gateway events are used for live state. Bounded REST reads are reserved for catch-up and verification.
Source Packets
Important messages become source refs with timestamp, content hash, attachment metadata, claim type, and confidence.
Read-Back
Posts and reactions require read-back before the system claims delivery or attention.
Boundaries
Private/no-agent-attention surfaces are excluded from live attention and only reviewed through explicit bounded requests.
Why It Matters
Customer-facing AI needs reliable surfaces: acknowledgements, attachments, delivery checks, and human-readable state.
- GatewayCapture live events rather than polling for state.
- Source PacketBind claims to source, freshness, and proof grade.
- ReactUse low-noise attention markers only after capture.
- Read BackVerify surface delivery with returned message metadata.
- ReticketMissing output becomes reusable pipeline work.
Voice, Vision, and Chess AI
Multimodal inputs routed through confidence and validators
I treat voice, images, board states, and attachments as first-class source surfaces. The system should not claim it understood audio or vision until it has metadata, bounded acquisition, transcription or captioning, language/content detection, confidence, and a route to the right action.
Voice
Attachment metadata, bounded download, hashing, STT, language detection, intent classification, and confidence reporting.
Vision
Image and screenshot handling as inspectable evidence, not decorative input.
Chess
LLM coach/judge workflows wrapped around validated game state, legal moves, replay, and learning trails.
Product Goal
Make multimodal AI useful for real users: explain, verify, recover from misses, and improve the workflow.
Why It Matters
Applied AI work becomes real when models handle messy human inputs without pretending confidence they do not have.
- AcquireCollect metadata and bounded source evidence.
- InterpretUse STT, language routing, visual inspection, or deterministic validators.
- RouteSelect response surface, model, verifier, and escalation path.
- RespondUse text or voice only when allowed and useful.
- EvaluateRecord confidence, misses, and improvement candidates.
AWS Secure Remote Relay
Low-cost access path with operational proof
I built a small AWS-backed relay pattern so a travel laptop could securely reach home lab machines without exposing management ports publicly. The same infrastructure path also hosts this portfolio behind TLS, HTTP security headers, and per-IP rate limiting.
Infrastructure
Minimal cloud footprint, infrastructure-as-code, DNS, TLS, web server hardening, and controlled ingress.
Operations
Connectivity tests, startup behavior, restart checks, reachability proof, and user-facing runbooks.
Security
Public web path separated from private management path, with rate limiting and security headers.
Support Fit
Combines AWS networking, customer-style troubleshooting, documentation, and proof-driven debugging.
Why It Matters
It shows the same operating discipline needed to help customers deploy AI systems securely and reliably.
- ProvisionCloud instance, DNS, firewall policy, and deploy scripts.
- SecureTLS, headers, rate limits, and bounded access assumptions.
- VerifyReachability, service status, headers, and browser rendering.
- OperateReadable runbooks and postproof for future recovery.