Five startups from PearX’s latest demo day are pitching AI for tasks ranging from changing 3D scenes to running models on-device and finding reusable industrial designs. The standouts point to a useful question for the next wave of AI businesses: can a focused tool beat a general-purpose chatbot where the real work happens?
Watch Desk analysis
What happened
TechCrunch spoke to investors about companies from Pear’s latest San Francisco demo day, held last week. Its five picks span spatial models, chip hardware, personal assistants, estate planning and industrial design. The companies’ descriptions and performance claims are pitches, not independent product evaluations.
Speridlabs says its Mundus model can modify parts of a 3D scene while keeping the geometry consistent. Saia says its in-development chip runs AI inference directly from flash storage, and plans to make test chips next year. Ren is building a personal assistant with user-set guardrails; Datum searches companies’ 3D design libraries by shape; Veros aims to use AI to recommend and manage trust structures.
Our top picks
- Mundus keeps the rest of the scene in view
Speridlabs says edits can change one part of a 3D scene without losing the geometry around it, a practical test for spatial models. - Saia is taking a different route to on-device inference
The startup says its chip runs AI from flash storage; test chips are planned for next year, so the performance claims remain to be tested. - Datum searches by shape, not just file name
Its Geometric Fingerprint technology is designed to find previously built parts across a company’s 3D design library. - Ren puts guardrails at the centre of its assistant
The startup says it keeps data on-device where possible or uses a private cloud, with actions checked against user-set limits. - Veros is aiming at a high-friction professional task
The company wants AI to help recommend trust structures and manage assets, work that currently involves attorneys and wealth managers.
Why it matters
These pitches target bottlenecks that broad AI tools do not automatically solve: keeping 3D geometry coherent, fitting inference into constrained hardware, finding the right design among thousands of files, or limiting what an assistant can do. If the products work as described, their value will depend less on a dazzling demo than on whether they fit into specialised workflows.
There are very different hurdles here. Saia still has to turn a chip design into hardware and substantiate its speed and power claims. Veros is seeking a trust charter, according to TechCrunch, while Ren’s privacy and security approach will need to work beyond a pitch. A demo-day shortlist is a signal of investor interest, not a stamp of technical or commercial proof.
Our read
The most compelling ideas start with a specific job rather than asking users to find one for AI. Mundus and Datum have clear tests: can they make useful edits without breaking a scene, and find the right part without a laborious search? Saia has the longer road from promising architecture to working silicon. Keep the claims in the “show us” column until products and evidence arrive; demos are charming, but customers tend to want the thing to work on Tuesday.
What to watch
- Whether Speridlabs publishes examples showing how well Mundus preserves geometry during edits.
- Saia’s test-chip progress and evidence for its speed, capacity and power claims.
- Whether Ren and Veros can demonstrate their stated safeguards and meet the requirements of their target markets.
- Whether Datum’s shape-based search saves engineers time in real design libraries.
Discussion spark: Which specialised AI tool here solves the clearest real-world bottleneck, and what evidence would convince you it works outside a demo?
Sources and evidence
- 5 startups that caught VCs’ attention at the latest PearX demo day – TechCrunch (5 October 2026, 17:26 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.