Blog
News and notes from AVKSoft.
-
Testing AI Agents With Independent Agents
A practical methodology for using Codex, Claude, OpenCode, and local LLMs as independent reviewers without mistaking repetition for validation.
-
Evidence-to-Claim Verification: Why LLM Reviewers Still Fail
A bounded GLM/Qwen review pilot with strict JSON schema validation showing why detecting a bad claim, classifying it, citing evidence, and correcting it are separate tasks.
-
When More Evidence Makes an AI Agent Less Reliable
What an Axum migration A/B investigation showed about retrieval redundancy, evidence saturation, and final-answer interpretation.
-
The Crash Window: Recovering AI Agent Actions Without Blind Retries
How Agent Loop preserves UNKNOWN across external-effect crashes, failed compaction, and session recovery without blind retries.
-
A Successful Search That Never Searched
What a real MCP differential test revealed about unknown source selectors, silent zero-hit results, and fail-closed retrieval diagnostics.
-
Building an AWStats-Style Log Analyzer in Rust
Why I built Stats, and what log rotation, SQLite checkpoints, native plugins, and a real release pipeline taught me about the less obvious parts of log analysis.
-
From Ready Pods to a Working System: Building a Distributed Kafka Chat Architecture
An experimental Rust and Kubernetes messaging architecture using Kafka HA, PostgreSQL, ZITADEL authentication, encrypted storage, and verified cross-gateway authorization.
-
Building a Cloud-Agnostic Kubernetes Platform for Multi-Tenant Applications
A reproducible RKE2 lab on KVM/libvirt, with Cilium WireGuard networking, Rancher, and per-volume encrypted Longhorn storage.
-
Why My AI Agent Stopped After 4–5 Iterations
A debugging story from Agent Loop: how I confused a conversation token budget with a per-request output limit, and why my local agent kept stopping early.
-
Building Agent Loop: A Local-LLM Agent Runtime in Python
Designing a Python agent runtime with local LLMs, MCP tools, configurable agents and skills, and reliable token-budget management.