Of the ten postings, Element Human is the most revealing — it's a small London startup quietly betting that computer vision applied to human attention is a viable commercial wedge. The posting is short, but every word is load-bearing.

1. The stack tells the story. Python, Postgres, PyTorch, GCP. No Kubernetes flex, no microservices alphabet soup, no Rust rewrites. PyTorch (not TensorFlow) signals a research-leaning team — PyTorch dominates academic ML and is the default when your engineers are reading papers, not just shipping inference endpoints. GCP over AWS is a quieter signal: GCP's TPU access and Vertex AI pipeline are attractive when you're training vision models on video data. Postgres as the only datastore mentioned suggests they haven't yet hit the scale where they need a separate feature store or vector DB — they're still in product-market-fit territory.

2. What they actually do. "Eye-tracking for attention, facial coding for engagement, implicit testing for memorability" — this is the ad-tech / market research stack reframed as ML infrastructure. They're competing with incumbents like Tobii and Affectiva, but going SaaS-platform instead of selling hardware. The phrase "we've just launched a new platform" is the tell: they pivoted from services or bespoke studies to a self-serve product, and now need engineers to scale it.

3. Skills and trends highlighted.

• Multi-modal CV — combining gaze, facial action units, and behavioral signals is harder than any one alone
• Webcam-based inference — eye-tracking historically required hardware; doing it in-browser is the unlock
• Privacy-sensitive ML — facial data under GDPR is a minefield; this is implicitly part of the job

4. Flags.

Green: Honest remote policy ("100% remote for the foreseeable future, probably until the end of summer" — this dates the post to early COVID, which is consistent with the 2020 thread). Specifying "a few days every quarter" is unusually concrete. The tight stack suggests engineering discipline, not resume-driven development.

Yellow: No mention of team size, funding, or seniority level. "More engineers" is vague — are they hiring two or twenty? The post got truncated mid-sentence on benefits, which is minor but suggests the founder wrote it themselves rather than running it past recruiting. No salary band (common for UK postings, but still a friction point).

Red: Facial-coding and attention-tracking products have a recurring credibility problem — the underlying science (Ekman-style emotion recognition) is contested. A buyer-side analyst would want to know how they validate accuracy claims.