HW Design-Space Agent

/about

Software and hardware

I'm Brendan Lynskey, a staff-level engineer with more than 25 years in hardware, and I write the software too. This project needs both: the agent, the tests and this site are software; knowing which architectures to offer the agent, what a cost model must count, and where a two-point calibration stops being trustworthy is hardware.

Links go to my personal projects on GitHub (and the sites built from them).

Software

  • LangGraph and LangChain agents

    This project: a LangGraph graph with Send fan-out, interrupt() for human approval, SQLite checkpointing and structured output whose schemas leave no field for a number, and a fleet of those graphs, orchestrated by code around a shared LangGraph Store. Also a LangGraph coding agent with sub-agents, MCP tools and a local Ollama provider.

    Personal projects:HW_Design_Space_AgentRaschka_Coding_Agent_LangGraph

  • Python simulators and models

    Deterministic, tested models that other code is checked against: the bit-exact CORDIC golden model and cost model here, an agent-loop simulator, and a SimPy discrete-event simulator of disaggregated LLM serving.

    Personal projects:Agent_Loop_SimDisaggregated_Inference_Sim

  • TypeScript and Next.js

    This site and a family of animated explainer sites, each driven by a TypeScript port that reproduces its Python reference exactly, rendered statically, with Web Workers for the heavy models.

    Personal projects:Agent Harnesses ExplainedAgent Protocols ExplainedAgent Context Explained

  • Testing and CI

    Parity tests between languages with no tolerance, frame tests on every animation, end-to-end tests at desktop and phone widths in light and dark mode, accessibility scans and Lighthouse budgets in CI; in the agent's CI, generated RTL checked against the golden model in two simulators, formal proofs, gate-level simulation and a synthesis smoke test.

    Personal projects:this site's CIthe agent's CI

  • MCP

    Hands-on Model Context Protocol playgrounds: tools, resources and prompts servers behind a gateway and an SSE bridge, each with a web client.

    Personal projects:MCP_Gateway_PlaygroundMCP_Supergateway_Playground

Hardware (25+ years)

  • FPGA and RTL

    SystemVerilog and VHDL, Xilinx certified, UltraScale+. Public examples: the synthesisable CORDIC modules this project starts from, and this project's generator of verified SystemVerilog for four CORDIC architectures (synthesised with Yosys, nextpnr-xilinx and Vivado 2025.2 for the FPGA, and with Yosys and OpenSTA on the open sky130 PDK); a parameterised AXI4 crossbar, a MESI cache controller, and an AHB-to-MII Ethernet MAC in VHDL with its testbench and verification plans.

    Personal projects:CORDICHW_Design_Space_AgentAXI4_CrossbarCache_Controller_MESIVHDL_example_code

  • SoC and RISC-V

    A RISC-V SoC integrating a 5-stage RV32IMC CPU, an AXI4 crossbar, MMUs, a MESI cache, DMA, an IOMMU and a PLIC, each verified on its own first.

    Personal projects:RISCV_SoC

  • Power: multiphase DC-DC and PDN

    Multiphase DC-DC conversion and power-delivery-network design. Public example: DC-DC converter control (PWM, PFM, hysteretic and constant on-time, with load-transient comparisons).

    Personal projects:DCDC_Control_Techniques

  • Signal and power integrity, thermal

    Seventeen interactive decks on signal integrity and high-speed digital design in which every number is computed by a model and checked against published work.

    Personal projects:Signal_Integrity

  • PCB: HDI, DDR, PCIe and MIPI

    High-density boards with DDR, PCIe and MIPI interfaces: schematic and layout practice, design for manufacture and board bring-up.

  • DSP (MSc)

    Signal processing, the subject of an MSc. Public example: a collection of browser-based DSP, synthesis and analysis projects.

    Personal projects:DSP_and_Music

Education

Contact

GitHub: BrendanJamesLynskeyLinkedIn

How it was built, and why

The project record is the engineering log: the principles every milestone follows, how the work is planned, done and independently verified with AI coding agents, every decision with its alternatives and reasons, what review found, and what is still missing. The final write-up is the repository's own closing report on all four milestones, every number in it bound to the committed data, and For a hardware team sets out, even-handedly, where it would help a hardware company and where it would not.

About this site

Every number here comes from HW_Design_Space_Agent at commit 80415c6 (535 recorded run traces: M1's 48 scored runs and one that crashed before scoring, M2's 60 runs and its 2-run pilot, M3's 110 structured, campaign, memory and pilot runs, M4's 314 (one per structured run and one per fleet member's run, pilots included); 3,253 LLM calls), vendored byte for byte with its hashes, or from the TypeScript ports of its golden model and cost model, which match the Python reference exactly. A Python script regenerates the site's data from the vendored files and checks every results row against the repository's own results table; the site's CI fails if anything drifts. Nothing calls a language model at runtime. The design and the animation framework come from the agent sites. Home · source.