iiinova — where your digital self keeps evolving

hev kit setup: brew install hev/tap/kit, set TURBOPUFFER_API_KEY, then hev up. That starts the layer gateway, a dashboard at 127.0.0.1:8099, and hevd indexing your coding-agent sessions. Needs macOS, Docker, and a turbopuffer key. hev.dev/kit/?utm_source=bluesky&utm…

System diagram titled 'What hev up starts', three groups left to right: your Mac running launchd contains session files from Claude Code and Codex feeding into hevd, the capture daemon (highlighted in orange), which indexes into the layer gateway running in Docker; the dashboard at 127.0.0.1:8099 also queries the gateway; the gateway archives into the hev-traces namespace in your own turbopuffer account. Caption: hev up starts all of it.

000

20ms baseline, 600ms after the reranker, flat relevance — that's the number that should gate the addition, not the conference-talk default. We hit a similar tradeoff benchmarking Jev as a reranker: real quality gain, but tail latency was the honest cost line.

000

Moving embedding into the query plan lets Turbopuffer overlap that call with other work. That is a concrete way to cut semantic-search latency without making embedding itself faster.

000

A real daily task, three samples per case, code checks plus a judge: this is a useful model comparison, and the Qwen 4-bit/6-bit reversal makes the memory and cache constraint visible.

Layer's Warehouse CRD names the upstream source a pipeline pulls from: Snowflake, Hugging Face Hub datasets, or any paginated JSON REST API, all shipped. Databricks and Iceberg are reserved in the schema, rejected by the operator until implemented. hevlayer.com/docs/kubernetes/wareho…

000

Rerankers now sit close enough in quality that nDCG on sparse, noisy human labels cannot tell them apart, Schmidt et al. argue. Their fix: LLM judgments calibrated across queries with item response theory, not just per-query relative comparisons. arxiv.org/abs/2609.35739

000

Mock embeddings that never got normalized to unit length, quietly skewing every cosine-similarity score until someone checked by hand, is exactly the bug class that survives a vendor demo and dies the moment you point it at your own repo.

000

Skipping the reranker only pays off if the seen-cache is actually doing dedup work, not just caching hits -- we measured the same tradeoff sizing Jev batch calls against a straight top-k cutoff, and it rarely comes free.

000

We built cost tracking into hev kit's trace archive: spend, tokens and cache hits by day and model at list rates, plus a heatmap of when you prompt. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: on your Mac, coding-agent session files flow to hevd, which indexes through the layer gateway in Docker. The gateway archives to the hev-traces namespace in your turbopuffer account. The highlighted Docker dashboard searches through the gateway and shows spend, tokens and cache hits, grouped by day and model at list rates.

001

Every proposal we send ships as an agent, not a deck -- one that has read your context and the proposed work, answers your team in our voice, and cites the real material. It stays on as the leave-behind after the engagement ends. hevmind.com

000

One traced session, 3.5 hours end to end: the agent spent 81% of it waiting on a person. That's the session that wrote hev up -- one data point kit made visible, not a general rate. hev.dev/kit/?utm_source=bluesky&utm…

Stat card, hev kit, one traced session. Headline: 81 percent. Label: of a 3.5-hour agent session spent waiting on a person. Subtext: one session, the one that wrote hev up, not a general rate.

000

Eval tooling forces an awkward choice: a heavyweight platform that is a commitment, or ad-hoc scripts that are fast but disposable. vibecheck is the middle path -- one YAML suite, runnable from a CLI a solo builder or small team can actually keep up with. hevmind.com/case-studies/vibecheck

000

Tying the compaction trigger to the prompt cache TTL instead of a fixed idle timeout is the detail that makes this efficient -- compact right before that 30-minute cache expiry, not after, so you're not paying to recompute a cache you were about to lose anyway.

Cohere is hiring a Partner Development Manager, Japan. We list the opening on our job board: jobs.hevmind.com/postings/cohere%3A…

000

Fusion matters in hybrid search. Target's authors compare merging strategies and choose weighted interleaving for lexical plus vector product search. They report roughly half as many zero-result searches as lexical-only search. arxiv.org/abs/2609.31498

000

Layer's Pipeline CRD lets extract/chunk and embed workers share one gateway queue through pipelineId. CPU extraction stages chunks; GPU embedding claims pending documents and writes vectors to the target VectorStore. hevlayer.com/docs/kubernetes/pipeli…

System diagram: a CPU worker extracts and chunks, then stages chunks in the Layer gateway queue labeled pipelineId: products. A GPU worker claims pending documents from the queue, embeds them, and writes vectors to the target VectorStore namespace.

000

Sliding-window LLM rerankers redo near-identical CoT reasoning per window, wasting latency. QReason reasons about the query once, reuses that chain across every window with a lightweight non-reasoning reranker -- matching or beating reasoning rerankers on BRIGHT. arxiv.org/abs/2609.30904

100

A trace replays in the harness's own shape -- your prompt, the agent's narration, tool calls folded until you open them. A rail down the side marks every prompt, failure, and wait. hev.dev/kit/?utm_source=bluesky&utm…

System diagram of hev kit. On your Mac: agent sessions from Claude Code or Codex feed hevd, the capture daemon, which reads them. In Docker: hevd indexes through the layer gateway; a dashboard (replay UI, highlighted) queries the gateway to replay sessions. The gateway archives to a hev-traces namespace in your own turbopuffer account.

000

Following one tool call end-to-end through the Aspire dashboard is what actually shows agent observability instead of just claiming it -- we built hev kit for the same instinct, pointed at coding-agent sessions after the fact.

000

Layer logs every query the gateway serves into a durable JSONL trail in S3, mirrored to a hot cache for recent reads. Fetch events tag back to it, so a search session is reconstructable after the fact -- not just the query, the whole interaction. hevlayer.com/docs/api/search-history

000

Turning the transcript archive into evidence for AGENTS.md edits is the clever part of this workflow; the incremental SQLite FTS import makes that evidence searchable. We use the same trace-as-corpus idea in hev kit.

000

Claude Code and Codex can search their own trace history through the hev-query skill: plan a few phrasings, fan them out via hev query, read the session around the best hits. No need to remember which session, just what happened. hev.dev/kit/?utm_source=bluesky&utm…

A stylized illustration of a diver's lamp shining a beam of light down through dark water onto a metal toolbox resting on the seafloor, surrounded by coral and rising bubbles.

000

Most retrieval pipelines make one pass over the corpus -- whatever the first candidate set misses is gone for good. Bigdeli et al.'s Seek fixes that without extra training: the retriever iterates against the corpus instead of committing after one shot. arxiv.org/abs/2609.28980

000

Type what you remember and kit finds where it turned up -- down to the command your agent ran. hev query "why did the preflight fail" searches prompts, replies, and tool calls across every session. hev.dev/kit/?utm_source=bluesky&utm…

A desk lamp shines an orange beam of light down through dark, murky water onto a glowing thread rising from the dark seafloor below.

000

Harness ownership deciding where traces and logs live is exactly the question kit is built around -- we archive outside the harness entirely, into your own turbopuffer namespace, so no single operator's choices decide what survives.

000

Six of seven harnesses deleting their own traces on request is a genuinely useful number -- it's the same reason we write hev kit's archive to its own turbopuffer namespace instead of trusting the session to keep its own record.

000

Every file your agent opened, every command it ran, every approach it abandoned -- gone the moment the session closes. hev kit indexes those traces with hybrid search so the history stays searchable. hev.dev/kit/?utm_source=bluesky&utm…

A sunken metal toolbox resting on a dark seafloor, with faint threads of light filtering down through the water from above.

000

580 pages scored for a few cents by batching Jev's fixed-format questions instead of paying for generation -- same shape we leaned on for reranking: one true/false call per candidate beats a chat completion once the output space is closed.

We hit the same tradeoff building a reranker on Jev: one true/false question per candidate beat general-LLM classification calls on latency and cost while matching quality -- exactly the shape closed-set routing and triage need.

4,000 Jev decisions for about $1.30 in a hackathon wildlife sim is a real data point on what per-decision judgment costs at that price point. We measured the same shape from the rerank side: batching drove Jev to $0.54 per 1,000 queries in our BEIR benchmark.

000

The error path is the detail worth noting: a program agent escalates control back to a parent LLM agent instead of crashing outright. We hit the same shape in layer's pipeline workers — long-running claim loops that only hand off when the deterministic path can't cope.

000

We checked before trusting it: re-running the batch call shifted scores 0.004 on average, 0.14 worst-case, never changing the top-1 result. A random shuffle of the shortlist scores 0.12, vs BM25's 0.667. hevmind.com/writing/jev-as-a-rerank…

000

We co-host a Maven lightning lesson, 'Three Patterns for Better Agentic Search,' with Doug Turnbull and Trey Grainger -- an hour on patterns that change agentic retrieval outcomes, not a vendor pitch. hevmind.com

000

The web app already handled prompt management and trace inspection. What broke the day-to-day loop was bouncing between IDE, terminal, web UI, and a separate cost dashboard for every check -- so we built a CLI to keep that loop in one place. hevmind.com/case-studies/evals-plat…

000

A 40-class emoji classifier trained overnight for about $10, running in RAM at ~150ms a query, is the same bet we made pointing Jev at reranking: the probability distribution is the output, not a side effect.

Cohere is hiring an IT Support Specialist. We're listing the opening on our job board: jobs.hevmind.com/postings/cohere%3A…

000

A hard 0.01 floor that never decays to zero is the kind of detail you need before thresholding on Jev's scores -- our own reranker leans on the low end of that range, where a floor like that would blur exactly what matters.

110