AnalogLab Harness

Inspiration

Analog IC design is iterative, evidence-heavy work. Students and junior designers move between requirements, hand calculations, operating points, and simulator results; when a specification fails, the hard part is usually understanding why and choosing the next safe change.

I wanted an engineering workspace that behaves like a senior analog mentor—not a chatbot, not an automatic circuit designer, and not a replacement for Cadence.

What it does

AnalogLab Harness turns a folded-cascode OTA goal and explicit specifications into an auditable workflow:

1. Compile a typed workflow plan.
2. Validate its tools, units, artifacts, risks, and bounded loop.
3. Require human approval before any engineering tool runs.
4. Run deterministic sizing, synthetic simulation, specification grading, and next-iteration proposal tools.
5. Preserve content-addressed artifacts and an append-only event trace.
6. Replay the exact saved decisions without calling a model or MCP tool again.

The public reference run intentionally fails gain and bandwidth once, explains the failed margins, applies a deterministic correction, and passes on iteration two. Engineers can also upload Spectre JSON/CSV results for deterministic normalization and grading; AnalogLab never launches Cadence.

How we built it

The product separates model reasoning from engineering computation. Python/FastAPI owns unit conversion, feasibility checks, saturation rules, spec margins, proposal rules, immutable artifacts, SQLite state, and replay. A Next.js/TypeScript workspace presents the target, plan review, approval boundary, run evidence, artifacts, and replay.

The workflow compiler is deliberately constrained: it receives only the goal, SI specification values, and allowlisted tool metadata. It cannot read PDK tables, netlists, historical runs, or raw simulation results. The code includes a strict OpenAI Responses API adapter for GPT-5.6 and a locally validated NVIDIA NIM compatibility path. The public sandbox defaults to server-managed NVIDIA NIM and labels that receipt truthfully; users may optionally supply an OpenAI or NVIDIA key for one HTTPS compile, and that credential is never stored in plans, events, artifacts, or evidence bundles.

Codex was our implementation partner for the contracts, deterministic tool migration, FastAPI/Next.js integration, test coverage, security boundaries, UI refinement, deployment, and the submission evidence package.

Challenges we ran into

The central challenge was avoiding a convincing but unverifiable AI engineering demo. We made the model compile a bounded workflow rather than calculate transistor sizes; the deterministic runtime produces every numerical result. We also added explicit approval gates, artifact typing, risk policy, bounded retries, failure states, and replay so a result can be inspected rather than merely trusted.

A second challenge was making a public demo honest: its simulator data is synthetic and visibly labeled, while imported Spectre data is treated as an uploaded artifact rather than proof that AnalogLab controlled a simulator.

Accomplishments that we're proud of

• A plan cannot execute MCP tools before approval.
• The same approved bundle replays with the same event sequence and decision digests.
• The public demo shows a real failed-specification iteration and a deterministic next proposal.
• Provider receipts expose the actual compiler mode instead of claiming that an NVIDIA call was GPT-5.6.
• The hosted sandbox works without an account and keeps anonymous sessions isolated and short-lived.

What we learned

Reliable AI engineering software needs explicit boundaries: the model should explain and structure decisions, while deterministic tools own calculations and checks. Auditability, clear failure states, and honest data provenance make the tool more useful for beginners than an unrestricted agent.

What's next

We plan to add topology packs for two-stage Miller and telescopic OTAs, approved Cadence/Spectre launcher MCP tools, parallel corner and Monte Carlo workflows, organization audit exports, and benchmarked proposal policies.

Built With

• codex
• docker
• fastapi
• gpt-5.6-responses-api-adapter
• model-context-protocol
• next.js
• nvidia-nim
• pydantic
• python
• react
• render
• sqlite
• tailwind-css
• typescript