Building AI systems that can remember, focus, speak, see, steer, delegate, recover—and make every token count.

I build local-first tools and infrastructure that make AI systems more useful, focused, and cost-efficient without making them more fragile—or taking continuity away from the people using them.

Explore the projects View GitHub profile

Selected open source

Five projects. One active experiment.

Practical infrastructure for attention, continuity, communication, and reliability—plus an open experiment in agent trajectory learning.

Cal

Local-first agent

Problem: Assistant continuity is often trapped inside a vendor's product boundary.
What it is: A personal agent harness with local memory, tools, routines, multi-channel access, and delegation to specialist agents.
Impact: The model can change while your continuity, history, and working context stay yours.

Featured Cal capability The Delegation Loop Cal frames work, delegates it to Codex, closes the question loop, and brings the finished result back. Explore Cal

Acton

Agent discovery

Problem: The agent ecosystem has no shared layer for cross-platform discovery, reputation, or intent-driven orchestration.
What it is: An action engine that discovers and ranks agents and tools, then orchestrates the best path to complete a goal.
Impact: 25,648 agents catalogued and more than 1.38 million reachability probes recorded since May 2026.

Explore Acton

Skiller

Context retrieval

Problem: Loading every skill into every request fills agent context with irrelevant instructions.
What it is: A dynamic retrieval layer that finds and injects only the skills relevant to the current intent.
Impact: 94% fewer skill tokens while maintaining outcome quality.

Explore Skiller

Talkbox

Voice orchestration

Problem: Adding voice often replaces the existing agent with a second system that lacks its memory and tools.
What it is: A voice harness that separates speech mechanics from the backend agent that owns truth and action.
Impact: Natural conversation without giving up the agent's brain, identity, or capabilities.

Explore Talkbox

AutoHeal

Agent reliability

Problem: Unattended agents repeat failure cascades until a person diagnoses and repairs the root cause.
What it is: Consent-gated causal diagnosis and repair with health checks, backups, and automatic rollback.
Impact: A reference deployment went from 196 errors a week to zero for 42+ consecutive days.

Explore AutoHeal

LAB

Loop Pilot

Active experiment

Research status · Phase 1: Observe

Problem: Agent loops either run without a fuel gauge or stop at rigid limits that ignore task complexity.
What it is: An open research experiment testing whether past agent trajectories can predict useful budgets, tools, and operational risk.
Impact: Phase 1 runs in shadow mode, collecting prediction-versus-actual evidence without changing live agent behavior.

Follow the experiment

The through line

Useful autonomy needs owned context and reliable boundaries.

These projects explore the layers around the model: what it remembers, what it sees, how it speaks, how long it works, and how it recovers.

Browse every repository Visit monikabishnoi.com