I got tired of watching AI agents make the same mistake twice with no way to know why.
So I built a diff engine. Then I needed to know when the facts it relied on had gone stale. So I built a staleness monitor.
Twelve libraries later, I had accidentally named the substrate layer of the agentic era.
Not frameworks. Not wrappers. The missing primitives — the things every serious agent stack eventually needs but nobody had bothered to build cleanly.
One sharp tool at a time.
Agents can act. Almost nothing yet verifies, tracks, or proves what they did. I am building that layer.
Creator of agentdelta — semantic diff engine for AI agent behavior. Find the exact decision fork between two runs. MIT-licensed, PyPI-published, MCP-ready.
Creator of foghorn — decision staleness alerts for AI agents. Warns when a fact an agent relied on has changed downstream. pip install foghorn-ai
Building 12 open-source Python libraries for the full agentic AI infrastructure stack — observability, attestation, computer-use primitives, multi-agent coordination, and game AI. All MIT-licensed · Austin, TX.
Open to partnerships and select projects — Building agentic AI infrastructure, MCP server design, agent observability. DM @sandeep_alluru
|
|
|
| agentdelta Semantic diff · behavioral replay |
foghorn Staleness alerts · agent memory |
groundcrew State oracle · computer-use |
|
|
|
| humanproof Bot detection · motor fingerprint |
agentcrdt CRDT · multi-agent state |
worldoracle NPC AI · belief repair |
 agentdelta Semantic diff engine — find the exact decision fork between two agent runs  
|
 foghorn Decision staleness alerts — warns when a fact an agent relied on has changed  
|
 groundcrew Deterministic state oracle and semantic action codec for computer-use agents  
|
 notarize Canonical trace format and tamper-proof verifier for agent execution attestation  
|
 clickproof Persistent GUI behavioral facts — what the agent saw and clicked, content-addressed  
|
 polaroid Embeddable CRDT scene graph — shared spatial memory for embodied AI agents  
|
 humanproof Motor-noise fingerprinting — distinguish human from bot by mouse trajectory alone  
|
 agentcrdt Semantic-causal CRDT — conflict-free agent-mutable world state with causal ordering  
|
 rulegraph Natural-language rulebook compiler — parse game rules into executable logic graphs  
|
 balancelab Economy red-team — Bellman-Ford arbitrage exploit detection for game economies  
|
 normsync World constitution engine — norm-governed behavior contracts for multi-agent games  
|
 worldoracle NPC contradiction detector and belief repair for game worlds  
|
| Research methodology | 97 ideas → 6-axis adversarial kill-test → 15 survivors → 12 shipped. Every library passed: novelty, defensibility, composability, market timing, implementation risk, and narrative fit. |
| Quality bar | 14-stage DevSecOps pipeline: lint · format · typecheck · tests (≥85% coverage) · SAST · secrets scan · dep audit · license check · build · pkg integrity · SBOM · release gate. All MIT-licensed, fully typed. |
| Stack | Python · MCP (Model Context Protocol) · PyPI · ruff · mypy · bandit · cyclonedx |
| Location | Austin, TX |
X · LinkedIn · YouTube · salluru.dev