Pax Hires SF Engineers for AI Trade Tower, No H-1B Sponsorship

The Control Tower Automates Duty Drawback

Pax, a ten-person Y Combinator company founded in 2024, is scaling an AI control tower that automates duty drawback and tariff recovery for importers, and it's hiring early-career software engineers in San Francisco without H-1B sponsorship. The platform already processes live CBP data for Verkada, Glossier, Yamaha, and Kodak, recovering millions in overpaid duties while the tariff regime compounds.

A customs entry is a legal assertion. Every field — value, origin, classification — carries the weight of a federal statute, and the penalty scales with the value of the goods, not the size of the mistake. For decades the only way to catch errors before they reached U.S. Customs and Border Protection was a human reviewer comparing spreadsheets against invoices, one line at a time. That model breaks when tariff schedules stack four deep and a single shipment touches a dozen Harmonized Tariff Schedule codes.

Pax built its platform on a different premise: the data already exists. Automated Commercial Environment records, entry summaries, commercial invoices, bills of lading, arrival notices: every document the supply chain produces lands in a silo. The company's AI control tower ingests those sources at the line level, reconciles them field by field, and runs a persistent audit across the full tariff stack: Section 301, Section 232, Section 122, and the IEEPA reciprocal rates that shift with each executive order. The system flags a Section 301 exclusion that should have zeroed a duty but didn't, an entered value that diverges from the invoice by twenty thousand dollars, a drawback opportunity buried in a year of filings. Each alert arrives with the expected value, the actual value, the source documents, and the reasoning, evidence a licensed broker can review and sign.

The regulatory boundary is explicit. In January 2026 CBP issued Headquarters Ruling H350722, its first decision addressing an AI-powered classification tool. The ruling drew a bright line: deciding what data appears on an entry constitutes customs business, and only a licensed broker may conduct customs business on another party's behalf. An automated tool is not a person under the broker regulations. A broker must specify the tool's decision logic and make the actual classification decision. Pax's architecture reflects that constraint. The platform drafts, researches, and narrows HTS candidates; the best current models return a fully correct ten-digit code roughly forty percent of the time without human review, but a licensed broker reviews every claim before filing. The company operates its own brokerage to close that loop.

Continuous screening replaces the periodic spot check. Traditional audits sample; Pax screens the full import program as data arrives. Pre-entry validation catches classification, valuation, and duty exposure before the entry hits CBP. Post-entry reconciliation matches filed entries against ACE and supporting documents, then routes prior disclosure, reconciliation, and protest holds. Post-export audit validates AES filings for value, license, HTSUS, and denied-party screening. A custom rule engine lets enterprises upload their audit playbook and convert it into enforceable logic that runs alongside baseline checks. Every decision (clear, hold, escalate) persists with actor, timestamp, and rationale.

The stakes are codified in 19 U.S.C. § 1592. A negligent false statement can trigger a penalty of two times the lost duties or twenty percent of dutiable value even when no duty was lost. Gross negligence quadruples that exposure. Fraud reaches the domestic value of the merchandise. An unchecked AI output does not lower the fault level. CBP's Advanced Trade Analytics Program now uses machine learning to build risk profiles from past transactions, examinations, and violations; consistency across filings keeps an importer off the target list. Pax's audit trail (evidence-linked checks, persisted timelines, cross-source reconciliation) is built to demonstrate reasonable care when CBP asks.

The platform reports estimated recoverable duties of $12.8 million across its customer base, with four drawback claims identified and a single Section 301 exclusion recovery of $21,270 documented in its product demonstration. Those figures represent money already paid: overpayments the importer had no practical way to surface before the control tower ran.

Tariff Surge Drives Demand

A single business bringing in containers of Chinese goods could face a six- or seven-figure annual tariff bill that was unfathomable a year earlier, paid at the border in cash long before any pass-through to customers.

The policy cascade moved fast. February 1, 2025: 25 percent on Canada and Mexico, 10 percent on China, justified by a fentanyl emergency under IEEPA. Within two months the scope expanded to nearly the full breadth of global commerce. April 2 — "Liberation Day" — the rationale shifted to reciprocal trade fairness. A 10 percent baseline duty on most imports took effect April 5, with deeper escalations for roughly 60 countries. Steel, aluminum, auto parts, and electronics saw rates of 50 percent or more. By July, a 25 percent duty on imports from India was announced for August 1, with no transition period.

Reciprocal tariffs operate on a different logic than the Most Favored Nation system that governed trade for decades. Instead of uniform WTO rates, they create country-specific rates based on bilateral trade imbalances and tariff disparities. The administration was already pivoting toward more durable frameworks, particularly Section 301 investigations spanning 76 separate tariff determinations, expected to conclude before the Section 122 surcharge expired July 24, 2026. Unlike IEEPA or Section 122, Section 301 tariffs carry no expiration date. If they materialize at the intended scale, they could approximate the breadth of the original IEEPA tariffs on far firmer legal ground.

For importers, the shifting rules hit the bottom line directly. Margins are tighter. Compliance is trickier. CBP enforcement has turned algorithmic. Automated tools now scan ACE and e-commerce channels for valuation anomalies, misclassification, and origin discrepancies. The Supreme Court's February 20, 2026 ruling in Learning Resources Inc. v. Trump — 6-3 that IEEPA does not authorize tariffs — triggered the CAPE refund process, but refunds are not automatic. Importers must file specific administrative claims. As of May 2026, claims have entered processing with billions disbursed.

The trade compliance software market reflects the pressure, growing at 12.3 percent CAGR. The structural complexity isn't receding — it's compounding. Companies that treat compliance as infrastructure rather than a cost center are the ones capturing refunds, avoiding penalties, and staying agile enough to restructure supply chains before options close.

Hiring Without H-1Bs: The San Francisco Filter

Pax's early-career software engineer posting drew over 200 applicants by late July 2026, with LinkedIn flagging it as "no longer accepting applications" within weeks. The role carries a base salary plus benefits (squarely in market range for entry-level full-stack positions), but the real filter appears in the fine print: "We work from San Francisco, in-person. Our office is focused, fast, and full of energy. This is a hard requirement, no remote." The company is hiring its first dedicated engineer outside the founding pair.

The visa stance is equally explicit. "Regarding visas: we will only support TN visas or transfers. Due to time, we will not sponsor new H-1Bs," the job listing states. That decision lands against a shifting federal backdrop: a September 21, 2025 rule added a $100,000 fee to new H-1B petitions for beneficiaries abroad for a 12-month window, and employers across the board are limiting sponsorship to high-impact roles while favoring visa-ready candidates. Tata Consultancy Services announced it would no longer hire through the H-1B program at all, pivoting to local U.S. talent. For Pax, a ten-person team with enterprise customers already on the books, the calculus is straightforward: speed beats sponsorship paperwork.

"We're well-capitalized, we already have real enterprise customers, and we're looking for an early-career software engineer to take real ownership of the products above, early enough that your work shapes the company and the $10B+ industry we're transforming," the listing reads. The pitch leans on ownership rather than brand prestige. Penny Chen, CEO, holds a PhD from MIT and built optimization algorithms at Flexport and on Amazon's supply chain last-mile team. Chris Le, the other founder, is a repeat founder with stints at Brex, TikTok, and Amazon's last-mile delivery optimization. Their argument: join now, write the core of an AI control tower that already generates 15% more refunds than the industry leader with 99% less processing time.

The strategy contrasts with Flexport, which added 13 roles in the past week alone, according to Zero G Talent's job board. The in-person mandate isn't unique to Pax; it's the prevailing mood among well-funded logistics and AI companies betting that density accelerates product velocity.

Pax's bet is that early-career engineers who can work in San Francisco without visa sponsorship represent an undervalued pool, especially as the administration argues the new H-1B restrictions will open more entry-level jobs and raise pay for American STEM graduates by reducing reliance on lower-wage H-1B hiring. The company doesn't need volume; it needs two or three engineers who can move fast on LLM-driven customs automation. The posting closed quickly. The next one will likely follow the same template.

The Founders' Edge: Logistics, Marketplace, Fintech

Pax's founding pair didn't stumble into trade compliance — they engineered their way into it from the inside. Penny Chen, co-founder and CEO, holds a PhD from MIT and spent years as a research scientist at Amazon and Flexport building pricing, forecasting, fulfillment, and capacity-planning algorithms for global logistics networks. At Flexport she saw the duty drawback machine up close: a program that returns billions in overpaid tariffs but only to enterprises large enough to absorb 10–20% broker fees and months of manual paperwork. Chris Le, co-founder and CTO, brings a second-time-founder track record across Amazon's supply chain scaling team, TikTok's zero-to-one e-commerce merchant systems in Singapore, and Brex's financial rewards and billing infrastructure. That combination (deep logistics optimization from Flexport and Amazon, consumer-scale marketplace plumbing from TikTok, fintech-grade money movement from Brex) is the specific vector Pax is betting on.

Pax started from the refund, not the shipment. Chen's Flexport tenure gave her the exact problem statement: 80% of eligible drawback (roughly $10 billion annually) goes unclaimed because legacy brokers won't touch claims under $100,000 and the manual process consumes months. Le's Brex and TikTok experience taught him how to turn regulated, high-volume financial workflows into self-serve software. The result is a system that ingests commercial invoices, bills of lading, and export proofs; extracts and validates line items; matches import-export pairs across thousands of permutations; calculates optimal refund scenarios; and files electronically, cutting processing time by roughly 99% and, per Pax's benchmarks, delivering 15% higher refunds than the leading incumbent software. The seven-person team hit $1 million in annualized booking revenue before its $4.5 million seed round led by Initialized Capital closed in April 2025.

That speed matters because the competitive moat in trade compliance isn't the model — it's the labeled data and the regulatory logic wrapped around it. Pax's bet is narrower and deeper: own the drawback calculation end-to-end, prove the 15% lift at SMB scale, then expand upward into bonded warehouse management, foreign trade zone planning, and export incentives. The founders' pedigree isn't a recruiting badge; it's the only reason they knew which regulatory edges to automate first.

Early Customers: Verkada, Glossier, Yamaha, Kodak

Pax's customer list reads like a cross-section of modern manufacturing and branded commerce: Verkada, Glossier, Yamaha, and Kodak all run their trade compliance on the platform. That statement comes directly from the company's own hiring materials, and the Verkada case study (published on Pax's site) provides the clearest window into what "running trade compliance on Pax" actually looks like in practice.

Verkada, an AI-driven physical security hardware company, imports components across multiple countries of origin. That supply chain geometry made it vulnerable when the U.S. tariff environment shifted in 2025: Section 301, Section 122, and IEEPA reciprocal tariffs stacked on top of each other, changing week to week. Manual duty-recovery approaches became impractical at the velocity the new rules demanded. Pax stepped in as what the company calls the "first AI-powered broker in the industry." The platform processed tens of thousands of PDF pages of entry documentation, extracted the underlying data, and centralized everything into a single system. It automated import-to-export matching, established ACE accounts and reporting infrastructure, and provided direct access to a licensed customs broker, enabling filings at scale. The result: Verkada has recovered significant tariff refunds, with claims already flowing back from CBP, and the two teams continue to identify new recovery opportunities as the tariff environment evolves.

Glossier, Yamaha, and Kodak appear on the same customer roster but with less public detail. Glossier's inclusion is notable given the company's recent financial restructuring; it secured a $45 million revolving credit facility in 2026 to fund its turnaround, suggesting that even capital-constrained brands see automated duty drawback as a recoverable cash flow lever. Yamaha and Kodak, both managing complex global manufacturing footprints, represent the industrial and legacy-manufacturing segments where customs error rates and unclaimed refunds have historically been highest. Pax's job postings frame the opportunity bluntly: each year, 80% of eligible import tax refunds (roughly $10 billion) go unclaimed, concentrated among importers under $50 million in revenue. The four named customers sit well above that threshold, signaling that Pax's initial traction extends into the enterprise tier where volume justifies the integration investment.

The validation matters because trade compliance software has long been a graveyard of pilot projects. Pax's early customer cohort suggests a wedge: companies with high SKU velocity, multi-origin sourcing, and exposure to the new reciprocal tariff regime need continuous, LLM-driven audit — not periodic batch processing. That the platform is already shipping entry and export audit agents to these accounts, per the product roadmap shared in hiring materials, indicates the control tower architecture is moving from demo to production inside real customs workflows. For engineers evaluating the company, the customer names are a proxy for technical credibility: the system is ingesting live CBP data, reconciling tens of millions in annual trade volume, and filing at scale, not simulating it in a sandbox.

The next hire will walk into an office where the tariff stack updates daily, the broker reviews every AI draft before filing, and the refunds are already hitting client bank accounts. The control tower doesn't run on slides. It runs on the entry line that just cleared.

Working in frontier tech? Zero G Talent tracks the openings: see every open Flexport role, browse frontier tech jobs, the companies hiring, and the people building the field.