The platform: autonomy, netlists, your method.
Everything an extraction needs — survey, decks, plan, fit, verify, report — in one engine you can drive at three levels of autonomy. Netlists are generated for you, libraries are updated surgically, and your own methodology becomes a playbook the engine follows every run.
Three levels of autonomy, one engine.
The same engine drives all three modes — hand it the keys, gate every step, or take the wheel yourself. Switch at any time, mid-extraction.
The AI does the modeling
Drop in measurements. The engine surveys the data, generates the decks, builds a corners-first extraction plan, fits stage by stage, self-verifies against quality gates and publishes the final report.
- Physics-ordered workflow: C-V → room-temp DC by effect → geometry scaling → temperature
- Every stage scored and accepted or retried — nothing silently passes
- Independent verifier AI on a second model cross-checks results
AI drives, you approve
Plan-first by design: the copilot proposes every stage — devices, diagrams, parameters, bounds — and nothing fits until you say GO. Intervene anywhere, then let it continue.
- Chat-native: “plan a full extraction, don’t fit until I say go”
- Review stages in the Extraction Cockpit before and after any run
- Disagreements between AI and verifier stay red until resolved
You drive, AI assists
Don’t want AI making modeling decisions? Use it as the fastest operator you’ve ever had: it does the software chores — setup, netlists, simulations, plots, reports — and touches nothing you didn’t ask for.
- Full manual cockpit: stage-based fitting, overlays, direct parameter control
- Cross-measurement diagrams: Vth vs L, gm_max vs T, custom expressions
- Model cards versioned with full lineage — every change traceable
You never write a netlist.
Every simulation deck AutoPDK runs, it writes itself — from the measurement set and the model registry. Traditional suites make you author test setups and netlists by hand: per device, per bias, per temperature, per simulator, and again for every C-V topology. Here that work simply doesn’t exist.
Automatic netlist generation
Instance lines, bias sources, sweeps, temperature, model loading, convergence aids and output extraction — synthesized for every device and every measurement, DC and C-V. Around twenty C-V topologies (cgg, cgd, cgs, cgb, cbd, cbs, cgc, cj…) are described declaratively and rendered on demand.
- Model registry is the single source of truth — instance naming, terminal order and OSDI module resolved automatically
- Change the model? Change the simulator? The decks regenerate. Nothing to rewrite.
- Existing Spectre, HSPICE and CDL model cards and libraries are read as-is
Surgical library updates
Point AutoPDK at a full PDK library — thousands of lines, dozens of models, corners and sections. It parses the whole thing, isolates the one model that needs re-extraction, updates it, and writes it back in place. Every other model, section and comment stays untouched.
- Target one model, one polarity, one corner — not the file
- Lineage recorded for the model that changed; nothing else is touched
Automate your own extraction methodology.
Nobody knows your devices like your team. AutoPDK doesn’t force a methodology on you — it learns yours. Describe your flow once, in plain language, and the engine executes it autonomously on every new technology. We’ve replayed a customer’s complete legacy extraction flow this way — stage for stage, green across the board.
Playbooks, not scripts
Your stage order, corner discipline, quality bars and naming — captured once as a plain-language playbook the engine follows every run. Your method, repeatable, at machine speed.
A sentence, not a setup
What takes an afternoon of macro programming in a traditional suite is one sentence here. The AI assistant does the software work; no proprietary scripting language to learn — or maintain.
You stay the authority
Using your own philosophy doesn’t mean trusting AI judgment. Gate every stage, review every fit, promote the metrics you care about — the engine optimizes against your definition of good.
Built for accountability.
Autonomy is only useful if it’s accountable. Four things make AutoPDK’s intelligence different from a chatbot with tools.
Show it a plot
Paste a screenshot — a bad fit from your previous tool, a curve from a paper — and the copilot sees it and acts on it. “Fix this region” is a valid instruction.
A team, not a bot
For hard judgment calls the copilot fans out parallel agents — independent critics review a fit, workers sweep alternatives, disagreement is surfaced, not averaged away.
An audit trail QA will sign
Every model card is versioned with full lineage: which stage changed which parameter, from which value, on which data. Reports link to the exact card generation they scored.
Agent-native by design
An MCP server exposes the whole platform as tools — connect your own AI agents, scripts or EDA flows and drive extractions, simulations and reports programmatically.
WAT / PCM wafer analysis
Parametric test data usually dies in spreadsheets. Drop a folder of WAT/PCM exports on AutoPDK — as they come from the tester, mixed lots and all — and get wafer maps, parameter trends and lot-to-lot comparisons immediately. Then ask for any plot you can describe.
Folder in, insight out
No format wrangling, no import wizards. The engine parses tester exports as-is, groups by lot, wafer, site and parameter — and untangles the messy folders too: mixed lots, split files, inconsistent naming.
Any plot you can say
“Median Vth per wafer across the lot — flag anything past 3σ.” That sentence is the plot spec: wafer maps, box plots, drift charts and site drill-downs on demand. Custom plotting without custom code — interactively, or through the API.
From fab floor to model card
PCM statistics live next to your compact models — one platform for monitoring the process and modeling the devices, with spec limits, gradients and outliers surfaced where decisions happen.
See the platform on your own data.
Early access teams get a working AutoPDK instance — cloud or on your hardware — and an engineer on the other end of the chat.