Aakash Gupta

Raytheon’s latest SPY-6 radar contract is worth $3.2 billion. One electrical engineer in Morocco just open-sourced a phased array radar you can build from Gerber files on GitHub. The cost ladder in radar is absurd. A Thales Ground Master 400 runs $30 million per unit. Morocco’s own air force bought eight Raytheon Sentinel radars for $67 million. The Navy’s SPY-6 engineering development contract alone was $386 million before a single production unit shipped. Commercial phased array systems for civilian use start around $250,000. The AERIS-10 does electronic beam steering at 10.5 GHz, pulse compression, Doppler processing, and multi-target tracking on a real-time map. The 20km version uses a 32x16 slotted waveguide array with GaN amplifiers, 16 ADTR1107 front-end chips, a custom frequency synthesizer, and an FPGA handling all signal processing. GPS and IMU for accurate target coordinates when the platform moves. This is a real radar system, not a science fair demo. The bill of materials for the extended version probably lands somewhere between $5,000 and $15,000 depending on component sourcing. Call it a 95% cost reduction from the cheapest commercial alternative. Everything is MIT licensed. Schematics, PCB layouts, FPGA code, Python GUI, all of it. The defense procurement complex charges what it charges because the technology was classified, the supply chains were locked, and the buyer had no alternative. Open source collapses all three of those barriers simultaneously. A university lab, a drone startup, or a national defense ministry in a country that can’t afford Raytheon pricing now has a starting point that would have required a cleared facility and a nine-figure budget five years ago. The creator is asking for beta testers, RF engineers, and FPGA developers. The project hit 20K views on X in 13 hours. That ratio of technical depth to viral speed tells you how much pent-up demand exists for radar technology outside the defense contractor paywall.

57% of merged PRs at Ramp in the last 24 hours came from a background agent. Most companies haven’t even started. The architecture Ramp built matters. Their agent Inspect runs in sandboxed VMs on Modal with full access to everything a Ramp engineer has: Sentry, Datadog, GitHub, CI/CD, feature flags, databases, live preview environments. The agent doesn’t just write code. It runs tests, checks telemetry, verifies frontend changes with screenshots, and opens PRs that pass the same review bar as human-written code. This is why the number is so high. Model intelligence was already sufficient. Environment parity was the missing piece. Once agents have the same context and tooling humans have, adoption compounds because engineers stop treating the agent as a side tool and start treating it as a parallel teammate running unlimited concurrent sessions. The product development implications are massive. PMs at Ramp now use Inspect during QA to make changes in real time instead of writing tickets. Designers can ship fixes without waiting for sprint capacity. The marginal cost of implementing a small change drops to near zero, which means the backlog starts to dissolve. Most engineering orgs are still debating whether to adopt Cursor or Copilot. Ramp already moved past the foreground agent phase entirely. Background agents that run autonomously, verify their own work, and produce merge-ready PRs at scale. The gap between companies measuring their agent PR ratio and companies that haven’t built the infrastructure to support one is going to define the next era of product velocity.

Anthropic PMs don't write PRDs. They use Claude Code to build the first version of the feature themselves. The whole company dogfoods it for weeks. Then they ship an experiment. Anthropic went from $1B run-rate revenue at the start of 2025 to over $9B by year end. Claude Code alone generates over $500M in run-rate revenue with usage growing 10x in three months. The company signed a term sheet at $350B, nearly doubling from $183B just months earlier. This is what happens when you eliminate handoff drift. In a traditional org, a PM writes a spec. Design interprets the spec. Engineering interprets the design. Each handoff introduces drift. Each handoff adds a week. The 2-4 week cycle of PRD → design → ticket → sprint → build → QA compresses to days when the PM can prototype in the codebase. The prerequisite is making your codebase AI-friendly. Tailwind instead of scattered CSS files. Microservices architecture. Reduced codebase size. That infrastructure investment is what lets PMs vibe code directly against production components. The math is simple. If your experimentation cycle is 8 weeks per test and your competitor's cycle is 3 days, they're running 13x more experiments per quarter. Compound that over a year and the learning gap becomes insurmountable. Walked through the full Claude Code project structure, skills setup, and how to get from zero to shipping in the talk.