The lab ledger, kept honestly.

Hypothesis

A site assistant can answer well and qualify leads in the same conversation, without gating basic answers behind contact forms.

Setup

DeepSeek streaming chat with a first-person persona, a 16k-token prefix-cached knowledge pack, full-text retrieval over the published writing, and a 12 req/min cap sitting under a 15/min edge rule. In the corner of every page, and inside the services, case-study and blog pages.

Current reading

Bilingual mirroring (Najdi in → Najdi out) is still the strongest trust signal. The newest reading is about punctuation: the prompt banned the em-dash as a machine-writing tell, and the model kept emitting one or two per reply anyway, because the prompt itself was full of them. Instructions ask, guards enforce. Measured cache hit rate runs above 99%, which is the only reason a pack this size is affordable.

Hypothesis

In a regulated topic, an assistant is only trustworthy if it answers from dated, approved sources. It has to refuse when those sources run out.

Setup

Grounded on hand-approved source chunks with dates, citation tags on every answer, a thumbs feedback loop into Supabase, and a 12 req/min cap.

Current reading

Refusing to improvise reads as competence, not weakness. Unanswered questions become the content backlog; the assistant tells us what to write next.

Hypothesis

Guided scoping beats a blank prompt box: a few sharp questions produce a better research prompt than one heroic attempt.

Setup

A strict-JSON dialogue: the model must ask one question per turn with clickable options, then hand back a copy-ready deep-research prompt in Arabic or English.

Current reading

Graduated to its own product page. The strict output contract (JSON only, validated) is what made it shippable. The constraint was the feature.

Hypothesis

Prompts written in the dialect people actually work in get copied and used more than prompts translated into textbook Arabic.

Setup

338 working prompts in Najdi Arabic across 7 categories (business, marketing, technology, education, design, government, productivity and wellness), with client-side search, filters, and one-tap copy.

Current reading

Dialect register is product infrastructure, not decoration. Now paired with a live Prompt Refiner that reworks rough prompts in the same register.

Hypothesis

A live simulator on realistic Saudi work scenarios sells an on-device Work OS better than screenshots ever could.

Setup

Pick a scenario and the model turns it into generated deliverables (reports, follow-ups, briefs) on safe demo data, capped at 15 req/min.

Current reading

People trust what they can poke. The simulator answers the "but would it work for OUR mess?" question before the first call.

Hypothesis

The first pass of a consulting engagement (name the bottleneck, order the moves) can run as a public tool without turning into theater.

Setup

One shared endpoint, strict JSON mode, vendor-neutral system prompts, validated outputs, 10 req/min. The advisor lives on the toolkit page, the refiner on the prompt bank.

Current reading

Fresh on the bench. Still watching what people actually paste in, and whether the "smallest move first" ordering survives contact with real workflows.

Hypothesis

Grounding a small model on real published content beats stuffing a bigger prompt. Second attempt: a model choosing from a list of titles beats cosine similarity over chunks, when the corpus is small enough to list.

Setup

Back, without embeddings. A cheap router call reads a manifest of every published document and names at most two; their full text is inlined into the answering call as a trailing system message, so message 0 stays byte-identical and the prefix cache holds. Retrieval failure is never fatal: it degrades to the ungrounded answer.

Current reading

The first version died with the single-provider decision, because DeepSeek has no embeddings API. The restart says the ranking was never the hard part: over 62 documents, an LLM picking from summaries is more accurate than the vector search that was deleted, and it costs a fraction of a cent. It also answers correctly in Najdi, and returns nothing at all for pricing questions, which is the behaviour that matters most.

Hypothesis

Sixty systematic homepage variants will find the brand’s visual voice faster than polishing one draft forever.

Setup

A private preview rig, reviewing structured homepage variations side by side: palettes, type pairs, hero compositions.

Current reading

Parked, learnings shipped: the warm-editorial system and the red "Thread" spine came out of this bench. The rig stays parked until the next big visual question.