Petrarch mines 700,000-plus bankruptcy records for physical AI

Petrarch Launches: A Marketplace for Real-World Operational Data

Petrarch launched in the Summer 2026 Y Combinator batch as a marketplace for enterprise operational data, revealing a bottleneck that has frontier robotics labs scraping bankruptcy courts for training material.

Three Harvard dropouts — Ian Lee, Samuel Hahn, and Sudhish Swain — founded the company. Their primary YC partner is Vivian Midha Shen. Lee previously served as a Fellow at Meridian AI building training data and GTM infrastructure for AI-native Excel in financial services and as a Technical Advisor at Scale AI working on quality filtering and annotation for agentic assistants and reasoning models. Hahn previously led GTM at Magier AI (Techstars '24), building agentic PII-redaction and data-compliance products, and worked as a venture analyst at Mainstreet sourcing pre-seed fintech and deeptech. Swain previously built computer vision software for medical equipment identification deployed in California hospitals as Head of Software Engineering at Clean Sweep Group and served as a research intern at Brigham and Women's Hospital analyzing spatial transcriptomic cancer datasets.

"We dropped out of Harvard to build the first market maker for enterprise data," the founders wrote in July.

Frontier labs have exhausted the public internet. The next generation of physical AI (robots, industrial agents, autonomous systems) needs structured operational sequences: tasks, skills, workflows, decisions, and outcomes captured inside real businesses. Almost none of that data has ever been for sale. Petrarch sources it from distressed companies, legacy industries, and bankruptcy proceedings, then anonymizes and restructures it before brokering it to applied AI startups and research labs.

The company has already transacted construction plans, autonomous driving recordings, and codebases. Zero Index has invested, though the amount remains undisclosed.

Petrarch frames the shift as moving from "farming" data (scraping, labeling) to "mining" it — extracting specialized knowledge embedded in economically vital businesses. The founders argue that training long-horizon enterprise agents requires data from real, revenue-generating operations, not synthetic approximations or hand-crafted expert demonstrations.

"AI agents are great at writing code, yet fail at basic office tasks because public data rarely captures how real businesses operate," the company states on its LinkedIn page. "Training on authentic enterprise workflows is essential for making AI dependable in real settings."

The marketplace targets manufacturers, construction firms, logistics operators, and automotive companies, starting with a mid-sized design-build firm holding 20 years of annotated floor plans, RFIs, and site footage, and a legacy auto manufacturer producing over two million cars a year with lidar and driving recordings. Petrarch anonymizes the data, structures it, and sells it to the labs that need it.

Why Public Data Starves Physical AI

The internet gave large language models a training corpus measured in trillions of words (machine-readable, pre-structured, and effectively free). Robotics has no equivalent. A manipulation policy needs paired observations and actions recorded during physical interaction, and that corpus does not exist at internet scale. The mismatch is structural: every robot trajectory must be physically executable, every action is tied to a particular body, and every failure can damage hardware, objects, or the environment. The result is a supervision gap measured in orders of magnitude.

Teleoperation remains the highest-fidelity source of that data, yet a skilled operator produces only a handful of clean episodes per hour. Quality degrades as fatigue sets in. Teams have typically sourced robot arms and consumer-grade USB cameras separately, then written custom driver code to bind them, producing datasets marked by low-resolution imagery, inconsistent calibration, motion blur, and timing drift across views. That mismatch between what models can absorb and what teams can capture is the industry's true chokepoint.

Physical AI training data is fundamentally different from text or image corpora. It is multimodal, time-synchronized, and captured from real or teleoperated physical interactions. It must contain synchronized RGB and depth video, LiDAR or stereo where relevant, tactile signals (pressure distribution, vibration, slip), force and torque readings at the contact point, proprioceptive data about gripper state, and often audio. A tactile spike at 1,500 Hz is meaningless without knowing what the vision stream and force sensor showed in the same millisecond.

Vision data tells the model what items look like. Nothing in standard training sets captures what they feel like, how they respond to force, or when a grip is about to fail.

Annotating this data is not image labeling with extra steps. Annotators must label grasp quality, slip moments, contact initiation and release, object pose inside the gripper, deformation under force, and temporal boundaries of sub-actions across synchronized sensor streams. Peer-reviewed manipulation benchmarks show that adding tactile data to vision-only training pipelines can lift manipulation success rates by roughly 20 percentage points, with another meaningful lift from joint visual-tactile pretraining. But every tactile data point requires a physical interaction — a robot or human actually touching, grasping, or handling something. That makes capture slow, expensive, and sensitive to rig calibration, so large-scale public datasets remain rare.

Simulation helps, especially for rare or dangerous scenarios. NVIDIA's Isaac Sim and Isaac Lab let Agility Robotics train its Digit humanoid safely. Google DeepMind's Gemini Robotics 1.5 and NVIDIA's Cosmos world foundation models push the boundary further. Yet sim-to-real gaps remain significant for contact dynamics, material compliance, and sensor noise. The strongest pipelines blend synthetic and real data rather than relying on either alone. A useful robot world model must preserve the variables that matter for action: 3D geometry, object permanence, contact, material properties, constraints, forces, and the consequences of robot motion.

The central bottleneck is not only policy learning. It is the absence of mechanisms that convert the world's abundant unstructured behavioral data (human motion, internet video, simulation rollouts, interactive demonstrations) into grounded robot supervision. Those sources contain rich information about tasks, goals, contacts, failures, and physical constraints, yet most of it is not directly usable by robot policies because it lacks embodiment-specific action labels, task semantics, and reward structure. A latent action is not a command. A progress signal is not necessarily a reward. A human strategy may not be executable by a robot.

This is why frontier labs are starving for structured operational sequences recorded in real industrial settings. The International Federation of Robotics recorded 542,000 industrial robot installations globally in 2024, lifting the operational stock to roughly 4.66 million units. Forecasts point toward 575,000 installations in 2025 and beyond 700,000 by 2028. Each deployment generates operational data that no public scrape can replicate. The labs that access it gain a corpus grounded in the physics, constraints, and decision logic of actual work, the very data Petrarch is building a marketplace to supply.

Bankruptcy Court: The Controversial Supply Chain

Petrarch's supply chain starts where most companies end: in bankruptcy court. The startup acquires proprietary codebases, project files, and payment records from Chapter 11 proceedings and distressed legacy firms, then clears and anonymizes the material before transfer.

The inventory is larger than outsiders assume. BankruptcyData.com, a commercial archive Petrarch's founders cite, maintains 600,000-plus records across 45 years of filings, with structured data on assets, liabilities, creditor matrices, and court dockets. The Federal Judicial Center's Integrated Bankruptcy Database adds another 100,000-plus Chapter 11 cases with filing dates, entity types, and basic financials. About 55 percent of those cases are dismissed or converted, leaving estates that still hold operational artifacts (CNC toolpaths, PLC ladder logic, maintenance logs, vendor payment histories) that never reached the public internet.

Petrarch targets the subset where those artifacts map to physical work: a metal-stamping shop's die-change sequences, a food-processing plant's sanitation workflows, a regional trucking firm's route-optimization spreadsheets. The company's Y Combinator page lists "codebases, project files, and payments" as the initial data categories. Payment records are especially valuable — they encode vendor relationships, lead times, and cost structures that pure sensor data cannot.

The legal architecture is thorny. Section 363 sales, the standard mechanism for selling assets free and clear of liens in bankruptcy, do not automatically extinguish third-party copyright claims on embedded software or licensed datasets. Spencer Fane's 2025 analysis of distressed AI-company bankruptcies warns that buyers "should not assume that bankruptcy court approval of the sale provides a litigation shield against copyright claims brought by third-party rights holders." Due diligence must reconstruct the model's training-data provenance and evaluate pending infringement risk against the estate. Petrarch's anonymization is designed to strip personally identifiable information and licensed third-party code before transfer, but the boundary between "operational know-how" and "protected expression" remains unsettled.

Courts are already policing AI use in bankruptcy filings. The Southern District of Texas Bankruptcy Court issues standing orders requiring disclosure of AI-generated content in submitted documents; other districts have followed. Trustees and the U.S. Trustee's office monitor for inaccuracies in AI-assisted schedules; errors become the debtor's liability, and potentially the attorney's. That regulatory climate shapes Petrarch's compliance stack: every dataset passes through a de-identification pipeline validated against current local rules and circuit-specific precedent, not a static model trained on stale statutes.

The economics favor the buyer. A mid-size bankruptcy firm handling a Chapter 11 with 500-plus creditors reported a 60 percent reduction in claims-reconciliation time using AI to match filed claims against the debtor's books. Petrarch argues the same class of structured operational data (tasks, skills, workflows, decisions, outcomes) is what frontier labs need to train robots that can actually operate in a factory, not just simulate one. The bankruptcy estate gets a new asset class to monetize; the lab gets grounded training sequences; the original engineers see their work repurposed without their consent. That last point is the friction the market has not yet priced.

Who Wins, Who Adapts

The marketplace Petrarch is building does more than match buyers with sellers. It rewrites the economics of how physical AI gets trained — and who pays the price when operational knowledge becomes a tradable asset.

Frontier labs are the immediate winners. NVIDIA researchers have described the current development loop as fundamentally reactive: deploy a model, wait for failures, patch the specific failure, repeat. "We are developing models faster and faster, and evaluation is still very far behind," a NVIDIA engineer said in a September 2026 technical talk. The bottleneck isn't compute or model architecture — it's coverage. Labs need examples that span the long tail of real-world conditions: the edge cases, the rare decisions, the workflows that never appear in internet scrapes. Petrarch's pitch (anonymized codebases, project files, payment records mapped to tasks, skills, and decisions) targets exactly that gap. When a lab buys a dataset of manufacturing workflows from a bankrupt automation integrator, it isn't buying more data. It's buying diversity. "Foundational models have uncovered that more data is not always better. We need diversity, proper covering of all situations," the researcher said.

The labs hiring most aggressively reflect this priority.

These aren't generic ML hires. They're data infrastructure roles built for the ingestion, cleaning, and evaluation of specialized operational corpora, exactly the pipeline Petrarch feeds.

Industrial operators face a sharper calculus. A factory that enters bankruptcy doesn't just liquidate machines; it liquidates the embedded knowledge of how those machines were actually run. The PLC logic, the maintenance logs, the shift-change handoffs, the payment trails that reveal which suppliers delivered on time — all of it becomes inventory. Petrarch's anonymization process means the original company never hands raw data to a third party. But the result is the same: proprietary operational sequences, stripped of identifiers, flow to labs building robots that may eventually compete with the operators who generated the data. The operators who adapt will treat their operational data as a balance-sheet asset, documenting workflows with future licensing in mind, structuring archives for clean extraction. The ones who don't will watch their hard-won process knowledge leave the building in a bankruptcy sale.

Engineers sit in the middle. The skill set is shifting from model-centric to data-centric. "We have converted the data problem, the data scarcity problem, into a compute, into a simulation issue," the NVIDIA researcher said — but synthetic data still regresses when real examples exist. "I'm not aware of any experiment that says pure synthetic data is better than real data." That means engineers who can design evaluation harnesses, build coverage maps, and trace disengagements back to root causes in operational data are now more valuable than engineers who can tune a loss function. The hiring bands bear this out: the premium goes to people who understand the structure of work (the tasks, decisions, and outcomes that never appear in internet scrapes) not just the structure of tensors.

The tension is structural. Labs need the long tail. Operators own it. Petrarch brokers the transfer. Whether the trade strengthens human work (as Petrarch's mission states) or accelerates its displacement depends on who writes the next generation of workflows, and whether the engineers building them have ever seen a real one. When a machinist's 20 years of tool-path adjustments become a training set sold to a robotics lab, the machinist is not a party to the transaction. The bankruptcy court extinguishes the employer's obligations; the data marketplace monetizes the residue. Whether that residue constitutes a trade secret, a collective work product, or merely exhaust is a question for the next litigation cycle — and for the unions, works councils, and legislatures watching the first deals close.

What This Story Leaves Out — and What Comes Next

Petrarch's marketplace sits at a specific intersection: distressed industrial firms sitting on proprietary operational records, and frontier labs building models that must operate in the physical world. That intersection excludes three adjacent conversations that dominate AI data discourse but follow different logic, different supply chains, and different legal regimes.

First, consumer LLM data. The web-scale text and image corpora that trained GPT-4, Claude, or Llama (Common Crawl, LAION, The Pile) are not what Petrarch brokers. Those datasets capture public expression: blog posts, Reddit threads, product reviews, Wikipedia. They do not capture how a CNC operator adjusts feed rates when tool wear shifts, or how a procurement team re-routes orders when a Tier-2 supplier misses a ship date. The "internet data" well is deep but shallow on operational semantics. Petrarch's founders have said explicitly that frontier labs "scraped the internet dry" and now need "authentic enterprise workflows": tasks, skills, decisions, outcomes mapped to the real structure of work. That is a different asset class.

Second, synthetic data generation. Companies like Gretel, Mostly AI, and Nvidia's Nemotron 3 Ultra pipeline produce statistically similar records without touching real customer data. That approach solves privacy and consent cleanly; it does not solve fidelity. A synthetic invoice distribution matches the mean and variance of real invoices. It does not capture the edge case where a purchasing manager overrode the ERP to expedite a line item because the plant floor called — and that override, repeated across 400 similar events, teaches a model the unwritten rule that "expedite" actually means "pull from safety stock." Petrarch's pitch is that the value lives in the exceptions, not the distribution. Synthetic data is a complement, not a substitute, and the market will price them differently.

Third, copyright lawsuits over web-scraping. The New York Times v. OpenAI, Getty v. Stability AI, and the class actions filed by authors and visual artists turn on whether training on publicly accessible creative work constitutes fair use. Petrarch's supply chain (bankruptcy court records, distressed-asset sales, anonymization of proprietary codebases and project files) operates under contract and insolvency law, not copyright's fair-use doctrine. The legal questions here are about asset transfer in Chapter 7 and Chapter 11 proceedings, data-processing agreements, and whether anonymization meets the "reasonable expectation of privacy" standard under CCPA, GDPR, and sector-specific rules like ITAR for defense-adjacent manufacturing data. Those are distinct dockets.

The open questions start with privacy. Petrarch says it clears and anonymizes data before it leaves a seller's environment. But "anonymization" has no single technical definition. K-anonymity, differential privacy, and synthetic re-creation each leave different re-identification risk surfaces. If a dataset contains payment records linked to vendor IDs that appear in public SEC filings, a motivated adversary can join them. The market has not settled on a certification standard (no SOC 2 for de-identification of industrial operational data) so buyers today rely on bilateral audits. That friction limits liquidity.

Regulatory compliance is the second frontier. Manufacturing data increasingly falls under export controls (EAR/ITAR), critical-infrastructure protection (CISA directives), and sectoral privacy laws (HIPAA for med-device makers, GLBA for financial-services vendors). A bankruptcy trustee selling a codebase may not know (or disclose) that the repo contains ITAR-controlled technical data. The buyer, a frontier lab training a generalist robotics model, may not have a compliance program to screen for it. Petrarch's role as intermediary could expose it to liability as a "broker" of controlled technical data under 22 CFR 120.16. No case law yet clarifies whether a data marketplace that anonymizes and resells inherits the transferor's licensing obligations.

The third question is market structure. Today Petrarch is effectively a two-sided matchmaker with a YC S26 badge. The supply side (distressed industrials) is finite and episodic; bankruptcy filings spike in cycles, not steadily. The demand side (frontier labs building physical AI) is concentrated: perhaps two dozen companies worldwide with the compute, talent, and product roadmap to ingest structured operational sequences. That concentration gives buyers pricing power. If the market matures, expect specialization: one marketplace for aerospace MRO logs, another for semiconductor fab recipes, a third for cold-chain logistics telemetry. Each vertical will develop its own schema, its own privacy regime, its own pricing conventions. Petrarch's bet is that the first mover who standardizes the ingestion pipeline ("tasks, skills, workflows, decisions, outcomes") becomes the de facto exchange. The bet is unproven.

The first datasets are moving (construction plans, driving recordings, codebases) out of bankruptcy courts and into the labs building the next generation of physical AI. The market is open; the ledger is still being written.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.