Thinking out loud on hard problems
- The Multi-Agent Coordination Tax: Mesh vs. Mediator
A fresh paper shows agent-to-agent messaging grows quadratically with team size, and enterprises are already eating the bill for it. Microservices solved this exact shape of problem years ago.
- SLI/SLO Observability, Built By Hand Before Any Vendor Touches It
I built a vendor-neutral SLI/SLO pipeline from scratch — OTel, Prometheus, Sloth, Alertmanager, Grafana — hit three real bugs along the way, and got a genuine burn-rate alert into my inbox. Here's the whole thing, screenshots included.
- Mentoring Team 21 to 3rd Place at RBC's TechPowHer Hackathon
How six co-op students went from a use-case brief to a working LangChain/LangGraph multi-agent build in a week — and took 3rd place out of 22 teams.
- MCP Is Giving Up Statefulness — and That's the Point
The Model Context Protocol's July 28th revision rips out session state, sampling, roots, and logging. It's not a cleanup pass — it's MCP admitting it was designed for a laptop and is now running in a data center.
- Self-Hosting Chatterbox on Apple Silicon: The Build Log, and the Bigger Bet Underneath It
A real build log for self-hosting Resemble AI's Chatterbox voice-cloning model on Apple Silicon — the judgment calls the docs skip — and why this pattern reaches well past TTS demos.
- The Agent That Forgets Everything: Cross-Session Memory Without Breaking the Audit Trail
The first five parts solved within-session problems. Part 6 addresses what happens when the context window closes — and why cross-session memory for a compliance agent is the hardest unsolved problem in this stack.
- 530,702 Tokens in the Index. 560 Reached the Model. All 12 Answers Correct.
Part 4 mapped the two retrieval architectures. Part 5 builds the experiment and runs the numbers — a k-sweep from top_k=20 to top_k=1 where accuracy never broke.
- Why Did the Wrong Tokens Arrive? Reranking vs. PageIndex
Parts 2 and 3 compressed the tokens that arrived. Part 4 asks the better question — why did the wrong tokens arrive at all? Two retrieval architectures answer it in opposite ways.
- 60–95% vs. 9.2%: What I Actually Measured With Headroom
The upstream claim is 60–95% token reduction. My measured result was 9.2% for the config I'd ship to a regulated client — and 25.6% for the one I wouldn't. Here's why that gap is the right story.
- 97% Fewer Tokens, Zero Model Changes: RTK at the Harness Boundary
The harness made the agent reliable — and expensive. RTK cuts the cost without touching the model, the prompt, or the pipeline logic. And it covers a lot more than git.
30 posts