Quick snapshot: what S25 smells like
S25 is saturated with agentic AI, developer tooling, and verticalized AI products for healthcare, finance, and robotics. Practically every demo day pitch now contains some blend of agents + automation + domain data. That means a lot of engineering-heavy problems—and a lot of interesting roles for folks who want to build infra, observability, and trustworthy AI systems.
My top picks (cool / fun / exciting for software people)
Below are the companies I dug into and why they caught my attention. Each entry explains what they build, the engineering hooks, and what I’d watch if I were hiring or thinking about integrating their product.
1) Embedder — "Cursor for embedded"
What it does: AI-first IDE and agent tooling for embedded / firmware engineers: generate drivers, run test/flash cycles, and speed up firmware development.
Why it’s cool: embedded development is still painfully manual—anything that automates driver generation, flashing, and test loops will save weeks per feature. This sits at an interesting intersection of low-level systems, CI for hardware, and agent orchestration.
Engineering hooks: hardware-in-the-loop CI, reproducible cross-compilation, test harnesses that can program physical devices, CLI agents that can flash, run tests, and debug.
What to watch: their CLI + agent integration (if it can actually run end-to-end workflows on real boards, this will be huge for prototyping).
2) Wedge — trust & monitoring layer for health AI
What it does: deployment and monitoring tooling that helps hospitals deploy, observe, and govern AI models safely.
Why it’s cool: real-world ML in hospitals is a nightmare—data drift, auditability, and regulatory needs create demand for robust ML ops and observability.
Engineering hooks: model monitoring, explainability pipelines, secure model serving, and compliance/audit trails. Lots of opportunity for feature flags, canarying, and latency-aware inference.
What to watch: integrations with EHR systems and their approach to privacy-preserving telemetry.
3) April — voice-first AI executive assistant
What it does: voice-powered assistant that manages email and calendar hands-free; live on the App Store.
Why it’s cool: voice-first UIs for productivity are an under-explored UX frontier. Building safe, accurate outbound-email agents (that can match tone and avoid hallucinations) is technically challenging and interesting.
Engineering hooks: NLU for calendar/email intent, secure Gmail + Calendar integrations, short-latency on-device components, and robust user-style personalization.
What to watch: quality of generated replies, failure modes (what happens when it gets it wrong), and privacy practices for email handling.
4) Prism — AI-native product analytics (session-replay agent)
What it does: watches session replays using vision models and surfaces friction points to product & engineering teams.
Why it’s cool: automating qualitative signals from session replay turns hours of manual review into queryable insights—great for small teams that can’t hire a dedicated UX researcher.
Engineering hooks: efficient video/session encoding, vision models tuned to UI, semantic search over user sessions, and integrations with existing observability stacks.
What to watch: model accuracy for detecting true friction vs intentional flows, and retention/privacy guarantees for replay data.
5) Plexe — build predictive ML models from a prompt
What it does: an agentic system that connects to your data, runs modeling experiments, and returns deployable models from a plain-language description.
Why it’s cool: if it works, it dramatically lowers the barrier to traditional ML workflows (feature engineering, model selection, backtesting, deployment).
Engineering hooks: safe data connectors, reproducible experiment pipelines, automated feature engineering, and model packaging for production APIs.
What to watch: how Plexe diagnoses failures and how transparent the generated pipelines are (the trust problem for auto-ML agents).
6) Mbodi AI — teach robots with language and demos
What it does: embodied AI platform that converts language + demos into reliable robot actions.
Why it’s cool: it’s a real fusion of perception, planning, and symbolic reasoning—applies to manufacturing and supply chain automation where scripting robots is still tedious.
Engineering hooks: sim-to-real transfer, robust perception stacks, action-parameterization, and safety checks for physical systems.
7) Aegis & Galen AI — two health AI startups worth scanning
Aegis: automates insurance-denial appeals and documentation; a very practical, revenue-recovery ML application.
Galen AI: personal healthcare agent that aggregates records and wearables; interesting for patient-facing ML and data integration.
Cross-cutting technical themes (why devs should care)
- Agent orchestration: many companies are stitching together multi-agent flows—one agent to fetch data, another to preprocess, another to call models, and another to evaluate. Building reliable orchestration primitives is a hot problem.
- Trust, observability, and safe-deploy: audit logging, human-in-the-loop checks, and rollback primitives are becoming standard parts of the stack.
- Hardware + software CI: embedded and robotics startups are trying to replicate software CI for physical devices—this is where a lot of smart engineering will happen.
- Low-latency & on-device components: voice agents and safety-critical integrations need predictable performance and hybrid on-device/cloud models.
Hiring / partnership signal — where to look for senior roles
- Platform engineers for model serving and observability
- SRE / infra with hardware CI experience
- ML infra (feature stores, experiment tracking, deployment) engineers
- Data privacy / compliance engineers for healthcare integrations
- Frontend engineers who can integrate UI session replay + model insights
Final takeaways
S25 is an engineer-forward batch. If you’re a software builder, there’s a lot to love: from firmwarish tooling (Embedder) to safety/observability (Wedge, Aegis), to human-facing agent UIs (April), to new ML productivity primitives (Plexe, Prism). These companies are building the scaffolding future teams will use to ship agentic AI and trustworthy systems.
Top comments (0)