<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title><![CDATA[Victor — Selected Works & Journal]]></title>
    <link>https://victorc.me</link>
    <description><![CDATA[Research, Autonomous Systems & Artificial Intelligence - Autonomous systems, machine cognition, and software engineering.]]></description>
    <language>en-us</language>
    <lastBuildDate>Sat, 12 Sep 2026 16:59:38 GMT</lastBuildDate>
    <atom:link href="https://victorc.me/feed.xml" rel="self" type="application/rss+xml"/>
    
    <item>
      <title><![CDATA[[Research] On the Economics of Context: Latency, Cost, and Cache Invariance in Frontier LLMs]]></title>
      <link>https://victorc.me/research/context-caching-cost-latency-analysis</link>
      <guid isPermaLink="true">https://victorc.me/research/context-caching-cost-latency-analysis</guid>
      <pubDate>Tue, 08 Sep 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[A systematic evaluation of KV-cache reuse on large prompt prefixes. Demonstrating how prompt structure and prefix invariance cut input token expenditure by up to 75% while reducing Time-To-First-Token from 1.4s down to 310ms. - Takeaways: Cache utilization yields maximum ROI on prompt prefixes exceeding 32,000 tokens evaluated more than four times within their TTL window. Time-to-first-token exhibits a 78% drop on 100k+ token codebases when memory states are preserved warm. Cache key invalidation: dynamic timestamps and volatile user headers must be isolated to prompt suffixes.]]></description>
      <category><![CDATA[Artificial Intelligence]]></category>
    </item>
    <item>
      <title><![CDATA[[Research] The Fallacy of Heavy Agent Frameworks: Returning to Deterministic State Machines]]></title>
      <link>https://victorc.me/research/why-most-agent-frameworks-overcomplicate</link>
      <guid isPermaLink="true">https://victorc.me/research/why-most-agent-frameworks-overcomplicate</guid>
      <pubDate>Wed, 26 Aug 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Why multi-layered autonomous agent abstractions frequently fail in high-stakes production systems, and how minimalist typed state loops deliver superior reliability, observability, and deterministic bounds. - Takeaways: Layered abstractions introduce opaque failure modes, hidden retry storms, and unmonitored token consumption. Strict JSON Schemas paired with an explicit while-loop provide deterministic execution and simple unit-testing guarantees. State persistence belongs in standard relational datastores or Redis key spaces, rather than serialized agent memory heaps.]]></description>
      <category><![CDATA[Autonomous Agents]]></category>
    </item>
    <item>
      <title><![CDATA[[Research] High-Dimensional Vector Search: Memory Geometry of HNSW vs Quantized Inverted Indices]]></title>
      <link>https://victorc.me/research/vector-indexing-hnsw-vs-ivfflat-memory</link>
      <guid isPermaLink="true">https://victorc.me/research/vector-indexing-hnsw-vs-ivfflat-memory</guid>
      <pubDate>Wed, 29 Jul 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[A deep examination of graph-based versus inverted-file vector indexing when scaling beyond 1,000,000 dense vectors. Architectural trade-offs between DRAM footprint, re-indexing pauses, and NDCG recall. - Takeaways: HNSW graphs require up to 1.5x additional RAM beyond raw vector data to store multi-layer bidirectional proximity graphs. Scalar quantization (FP16 to INT8) paired with Matryoshka dimension truncation preserves 98.8% recall while decreasing memory by 65%. In memory-constrained environments, Inverted File (IVFFlat) structures with cross-encoder re-ranking deliver optimal cost efficiency.]]></description>
      <category><![CDATA[Performance]]></category>
    </item>
    
    <item>
      <title><![CDATA[[Journal] watchduck: Orchestrating MCP and agent concierges for Claude Code]]></title>
      <link>https://victorc.me/notes/watchduck-agent-concierge</link>
      <guid isPermaLink="true">https://victorc.me/notes/watchduck-agent-concierge</guid>
      <pubDate>Thu, 10 Sep 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[When using Claude Code and autonomous agents across multiple workspaces, managing tool registration and context boundaries requires an active concierge. Designed watchduck to handle protocol dispatch, session monitoring, and real-time execution feedback seamlessly.]]></description>
      <category><![CDATA[Architecture]]></category>
    </item>
    <item>
      <title><![CDATA[[Journal] lazyduck: Detecting and curtailing invisible AI token over-usage]]></title>
      <link>https://victorc.me/notes/lazyduck-token-overusage</link>
      <guid isPermaLink="true">https://victorc.me/notes/lazyduck-token-overusage</guid>
      <pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Agent loops often suffer from silent context accumulation: redundant tool definitions, repeated file dumps, and runaway self-corrections. Built lazyduck to observe token growth trajectories in real time, alerting and throttling before costs compound exponentially.]]></description>
      <category><![CDATA[Observation]]></category>
    </item>
    <item>
      <title><![CDATA[[Journal] ECC: Performance optimization harness for frontier coding agents]]></title>
      <link>https://victorc.me/notes/ecc-agent-optimization-system</link>
      <guid isPermaLink="true">https://victorc.me/notes/ecc-agent-optimization-system</guid>
      <pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Research-first development harness optimizing agent performance across Claude Code, Codex, Opencode, and Cursor. Focuses on instinct learning, durable memory stores, and sandboxed security harnesses for production coding tasks.]]></description>
      <category><![CDATA[Architecture]]></category>
    </item>
    <item>
      <title><![CDATA[[Journal] worldmonitor: Situational awareness through real-time AI news synthesis]]></title>
      <link>https://victorc.me/notes/worldmonitor-intelligence-dashboard</link>
      <guid isPermaLink="true">https://victorc.me/notes/worldmonitor-intelligence-dashboard</guid>
      <pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate>
      <description><![CDATA[Unified intelligence dashboard tracking geopolitical shifts, infrastructure alerts, and breaking technology events. Combines continuous web ingestion pipelines with LLM narrative clustering to isolate signal from ambient noise.]]></description>
      <category><![CDATA[Dispatches]]></category>
    </item>
  </channel>
</rss>