iiinova — where your digital self keeps evolving

hev query takes what you remember, not an exact string. Describe a failure and it surfaces the command your agent actually ran, searchable again after the terminal closes. hev.dev/kit/?utm_source=bluesky&utm…

000

Routing inktober submissions through clef-flash's image classification instead of a heavier VLM fits the feed-generator use case well -- constrained categorization at build speed, not chat latency. We saw the same payoff running Jev as a reranker instead of a general chat model.

000

We ranked 33 systems with a three-corpus mean. Mixedbread v3.1 listwise leads at 0.511 nDCG@10; Jev pair reaches 0.502 and Perplexity choice 0.501. The builder lets you cut by price instead. Free account: hevmind.com/research/rerankers/?utm…

Ranked plot of 33 rerankers, decision models, LLMs and retrieval baselines by mean nDCG@10 across SciFact, NFCorpus and FiQA. Mixedbread v3.1 listwise leads at 0.511; Jev pair is 0.502 and Perplexity choice is 0.501. The builder link and free-account note appear below the chart.

000

The discarded route in a coding-agent session can matter later: files opened, commands run, and approaches abandoned. We built kit to keep that trace searchable after the session closes. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: On your Mac, hevd reads coding-agent session files. It indexes through the layer gateway in Docker, which archives traces to the hev-traces namespace in your turbopuffer account. The trace includes files opened, commands run, and approaches abandoned.

000

In our reranker test, Clef-flash scored 0.498 mean nDCG@10 as 30 pair calls. As one batch call, it scored 0.283, below the BM25 order at 0.404. Request shape mattered. hevmind.com/writing/jev-has-company…

000

The network trace is the useful artifact here: it exposes the semantic side that the UI hides while still showing the Boolean expansion. That gives searchers a way to audit both retrieval legs.

000

We plotted reranker quality against price per 1,000 queries across SciFact, NFCorpus and FiQA. The chart builder holds 51 systems; you can change the axes or isolate one corpus. Free account: hevmind.com/research/rerankers/?utm…

Scatter of mean nDCG@10 against price per 1,000 queries for 18 rerankers and decision models. Perplexity choice, Jev batch and Voyage rerank-3 cluster near 0.50 at about fifty cents; Mixedbread v3.1 listwise is highest at 0.511; Perplexity batch, Liquid d1 and Tev1 sit above $10.

000

Benchmarking RAG against the standalone model caught negative lift, and replacing fixed-token chunks with atomic facts addresses the forum-thread failure at its source.

000

Fractional CTO work earns trust the way code review does: by reading what's actually there. We check a vendor's claims against the docs, a hire's answer against what the role needs, and an architecture against the roadmap it has to carry. hevmind.com

000

Finding the single-word entity cases that slipped through a green relevance check is a useful failure analysis. The API was healthy while retrieval sent the answer to the wrong country.

000

Follow-up to Jev meets Jev: added Perplexity's pplx-decider and Cloudflare's Clef. Perplexity edges Jev on mean nDCG@10 (0.5015 vs 0.5009) at $0.40/1k vs $0.54. Most gaps aren't statistically distinguishable. hevmind.com/writing/jev-has-company…

Horizontal bar chart titled Reranker cost per 1,000 queries. Four bars, cheapest to most expensive: Perplexity decider $0.40, Jev batch $0.54 (highlighted in orange), Clef-flash pair $1.39, Mixedbread v3.1 $1.42.

000

Collapsing logs, traces, analytics, alerts, dashboards, and telemetry export behind one SQL-queryable API is the right fix for the tab-switching that eats most observability debugging time.

000

hevd indexes every coding-agent session on the Mac as it grows. A trace can contain secrets or customer data. We do not redact or filter by project in kit yet, so run it only where you are comfortable archiving every session. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: on your Mac, hevd reads every coding-agent session file and sends traces to the layer gateway in Docker. The gateway stores them in a hev-traces namespace in your turbopuffer account. A note says kit does not redact or filter by project yet.

000

Cohere is hiring a Solutions Architect, Defence, DACH in Germany. We list the opening on our job board: jobs.hevmind.com/postings/cohere%3A…

000

MRVQ’s post-hoc code gives frozen embeddings two serving knobs, rate and dimension, while keeping one resident index. Avoiding a separate quantizer for each rate is the useful systems choice here.

000

We store kit’s trace archive in a namespace in your turbopuffer account. hevd reads local session files and indexes through the layer gateway. Transcripts can hold secrets; kit does not redact or filter by project yet. hev.dev/kit/?utm_source=bluesky&utm…

000

Zenovkin and Björkqvist train a neural patent parser on 1 million rule-parsed documents. They report better citation recall than its teacher at 3× lower inference cost. We find the local-attention design useful: train on short sequences, run beyond 40,000 tokens. arxiv.org/abs/2610.01553

000

We expose two signals in Layer’s namespace metadata: is_stable reports the latest index-status poll; stable_as_of keeps the timestamp of the last stable poll. On cold start the timestamp is null. hevlayer.com/docs/api/namespace-met…

System diagram: a serving namespace’s index status is observed by the Layer gateway’s namespace watcher. The watcher updates the metadata API, which returns is_stable and stable_as_of to the client. On cold start, is_stable is false and stable_as_of is null.

000

Using measured agent-hour cost as a census proxy is a sharp method: ~$16-18/hr for Codex against ~$24-50 for Claude Code, from real sessions rather than list-price guesses. We track that same actual-spend number per trace in kit, by model and by day.

000

We read Yun et al.'s CANOPY as two separate jobs: choose the right regions within retrieved text, tables, images and video, then retrieve again when the evidence is thin. Their ablations say the extra retrieval drives the multi-hop QA gains. arxiv.org/abs/2610.00923

System diagram of CANOPY: retrieved text, tables, images and video become item hierarchies; query-scored regions are selected; an evidence critic checks them and can request targeted retrieval. Newly retrieved items are compressed before joining the evidence.

000

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

000

hev kit is Apache-2.0: github.com/hev/kit carries the source, the CLI commands, config, and the design RFCs behind how it indexes and searches your coding-agent traces.

000

On macOS, brew install hev/tap/kit, set TURBOPUFFER_API_KEY, then run hev up. We start hevd under launchd, the layer gateway and dashboard in Docker, and archive agent traces in your turbopuffer account. Docker must be running. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: coding-agent session files on your Mac feed hevd under launchd. hevd indexes through a layer gateway in Docker, which stores traces in the hev-traces namespace of your turbopuffer account. A Docker dashboard at 127.0.0.1:8099 searches through the gateway. macOS, Docker, and a turbopuffer API key are required.

000

Training the tokenizer, generator, and reranker jointly instead of bolting a reranker onto a frozen codebook goes straight at codebook collapse and item collisions — the failure modes that usually get patched after the fact, not designed out.

000

Choosing a vector database by feature checklist misses the real decision: what search system you already run, what cloud credit you have left, and how fast your data is actually growing. Pricing calculators answer the wrong question. hevmind.com/writing/how-to-choose-a…

000

Turbopuffer is hiring a database engineer in the US / Canada. We have the role and its application link on the job board: jobs.hevmind.com/postings/turbopuff…

000

Jev reranker, end to end, is about 90 lines: the prompt, a request schema, chunking, a prune threshold, and the confidence-interval math behind the results. github.com/hev/reranker, pip install hev-rerank.

000

Cache hits change how to read a coding-agent cost chart. We show spend, tokens, and cache hits by day and model in hev kit, at list rates from the archived traces. hev.dev/kit/?utm_source=bluesky&utm…

000

We built kit to split coding-agent wall time among the person, the model, and tools, with context size underneath. A long session can mean very different things depending on where it waited. hev.dev/kit/?utm_source=bluesky&utm…

000

Kim et al. build a retrieval benchmark from scientists' own citations: papers they say actually advanced a project they completed, used as ground truth for which prior idea a new problem needs. AI search still lags that judgment. arxiv.org/abs/2610.02202

000

Tracing an apparent RAG regression back to exact-match grading of fuzzy answers is a useful eval fix: it stops a scoring artifact from masquerading as a model failure.

010

shelf.hevlayer.com puts one search box in front of three routes. Type an author, title, or vibe and Layer's Auto rank expression routes it to keyword, semantic, or a fused blend. hevlayer.com/docs/demos?utm_source=…

Diagram: a search box on shelf.hevlayer.com labeled 'author, title, or vibe' sends an arrow labeled 'query shape' to a highlighted orange box in the layer gateway labeled 'Auto rank expression, keyword · semantic · fused blend'. Caption: Live: shelf.hevlayer.com · github.com/hev/shelf.

000

Diffing the result count between plain search and EBSCO's new "AI-assisted search" -- both land on 1,006 -- while the order differs pins the feature on the reranker, not retrieval. Cleaner diagnostic than parsing the vendor's description.

000

Zhang et al. train a coding agent to decide when to compact context, not just react to overflow: a judge reviews its compaction calls, then RL jointly optimizes coding and compaction. Pass rates rose 9.2 points on SWE-bench Verified, 5.0 on SWE-PolyBench Verified. arxiv.org/abs/2610.02163

000

The notebook-search example makes the failure concrete: an HNSW scan over every team can leave too few candidates once the team filter runs. That is a useful check on any shared retrieval index.

000

We built kit replay for the part a search hit leaves out: the prompt, agent narration, and folded tool calls in the order they ran. A rail marks failures and waits so you can inspect what happened around the hit. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: On your Mac, session files feed hevd. hevd indexes through the layer gateway in Docker, which archives traces to the hev-traces namespace in your turbopuffer account. The Docker dashboard searches through the gateway and replays traces; a rail marks failures and waits.

000

Agent-directed doesn't mean no review -- it moves where review happens. Before: read every diff. Now: decide what ships without one, what pages a human, and when a passing test still isn't enough. We help CTOs draw that line on purpose, not by default. hevmind.com

000

We ship a hev-query skill for Claude Code and Codex. The agent plans several phrasings, fans them out through hev query, and reads the session around the best hits. hev.dev/kit/?utm_source=bluesky&utm…

System diagram: on your Mac, the hev-query skill for Claude Code and Codex sends hev query searches to the layer gateway in Docker. The gateway queries the hev-traces namespace in your turbopuffer account. The skill tries several phrasings and reads the session around the best hits.

000