Posts

Engineering ·
Reducing LLM Calls and Fixing the Cache and Morning Batch
With 1.6 days of credits left, how regeneration, reasoning tokens and call counts were cut, and how a cache that behaved differently from what we intended got fixed.
AI Agents ·
Keeping an LLM from Making Up Investment Numbers
In FINCH the LLM writes placeholders instead of numbers, and the server fills in engine values. Why it was designed this way, and three holes that showed up in production.
Engineering ·
Finding What Vector Search Missed with Keyword Search
When the documents grew 4.5 times, vector search alone dropped to 0.375 Recall@5. This post covers Korean bigram keyword search, RRF, and title, source and freshness weights, alongside stage-by-stage results and their comparison limits.
Engineering ·
When Output Checks Threw Away Good AI Answers
FINCH runs 10 automatic checks before sending an LLM explanation. Counting rejections showed many were false positives, mostly caused by the prompt and the checks disagreeing, not by the model.
Retrospective ·
Development Experience vs. AI Experience, Seen While Training New Developers
Helping with AI training for junior developers led me to reflect on hands-on experience, the responsibility to check results, and connecting AI use with people’s work.
Retrospective ·
When Code Can't Leave the Building: Janus, a Local LLM Agent IDE
Seeing teams that can't use external AI for security reasons, I tried running a coding agent on a local 27B model. A record of how much a harness could make up for a weaker model.
Engineering ·
Why Janus Stopped Giving Agents a Separate Worktree
Janus first gave every Task its own git worktree. To see results right away it now works in the original checkout. How the risks git can't undo were documented and contained.
Engineering ·
Why Janus Runs Three Concurrent Generations, Not Five
On a 48GB Mac running a 27B model, concurrent generation requests went from 5 to 3. The reasoning on KV cache, a single GPU and MTP, and the measured condition for raising it again.
Retrospective ·
Agreeing on How We Develop Before Bringing AI Agents into the Team
Applying AlgoSu’s AI agent practices to PinLog started with aligning people: making assumptions explicit and agreeing on structure, conventions, and ownership.
Engineering ·
Runtime Contracts and Infrastructure Ownership Boundaries for Deployment
Runtime contracts, infrastructure ownership, and the gap between passing checks and deploying: a look back at PinLog’s July 2026 records.
Engineering ·
PVC Retention Policies and Data Recovery Limits on Single-Node k3s
Looking at local-path and local-path-retain on a single-node k3s cluster, I separated three questions: Pod restarts, PVC deletion, and recovery after losing the node.
AI Agents ·
Separating an Operations AI’s Memory from Actual Server State
Separating observations, contracts, evidence, and query views when using Pico’s memory for operational decisions.