$246,250 Average AI Engineer Salary in San Francisco Surges

The Hiring Signal

Hera AI posted for an Industrial Software Engineer in the San Francisco Bay Area: entry‑level, full‑time, on‑site or remote. The hire will design, build, and maintain the software systems powering Hera's AI platform for manufacturing and engineering workflows. A separate listing for a Founding AI Engineer at Hera (YC S25) described the mission as building "the world's first AI motion designer that creates professional animations in seconds," a different company from the manufacturing intelligence startup.

The posting signals a broader surge. Built In SF reported in September 2026 that San Francisco's AI sector is surging, with average base salaries for AI engineers in the Bay Area hitting $246,250. The surge is driven by established leaders and rapidly scaling startups pivoting from experimentation to delivering large‑scale, agentic enterprise solutions. AI engineering has matured into its own discipline, distinct from traditional data science or machine learning roles. Many firms now require at least part‑time office presence, further intensifying talent migration toward the city's urban core. San Francisco has become a global AI headquarters, with artificial intelligence companies leading an office‑leasing resurgence to its highest level since before the pandemic.

Hera's own trajectory adds context. Meera Patel and Noelle So founded the company in 2026. Patel previously delivered projects for Pfizer and TE Connectivity, built factory automation systems used by millions, and published ML research at the University of Washington. So built water infrastructure at 19 serving companies behind 6% of Philippine GDP, developed software for engineering facilities at TE Connectivity and Wyze Labs, and was a former machine learning and industrial engineering researcher at UW. They met on a bus to a factory tour.

That background shaped the thesis: physical production spends too much time fixing misunderstandings about what was supposed to be made. The startup launched publicly as part of Y Combinator's Summer 2026 batch with two employees in San Francisco.

The numbers show a company moving from a two‑person founding team to active hiring for specialized industrial software roles in San Francisco at the moment the broader market for that talent peaks. Those hires are translating into an automated CAD/QA platform that has drawn aerospace attention, though not the collaboration the name might suggest.

From Hires to Hardware: The Drawing‑Review Tool

Hera launched publicly via the batch. "Today we're launching Hera to help back that engineer up," the company announced on LinkedIn. The "engineer" is the senior engineer who, in every custom‑metal shop, sits at the end of the line with a red pen and a stack of 2D drawings — the last line of defense before metal gets cut.

The product didn't emerge from a vacuum. Patel and So ran more than 200 conversations with manufacturing engineers, quality engineers, and machinists before writing a line of code. "The same story kept coming up: the last line of defense against a bad drawing is one senior engineer with too much on their desk," Patel wrote in the launch post. So grew up in an industrial park her family runs in the Philippines and at 19 built water infrastructure serving them. Patel taught herself PLC programming and C++, published machine‑learning research there, and built factory automation systems at TE Connectivity.

That customer discovery shaped the architecture. Hera doesn't try to replace the engineer's sign‑off. "AI reads the documents. Math checks the answers. It doesn't sign the drawing. The engineer does." The platform runs four automated checks, each exposed as a named module:

The checks cover GD&T conformance, code compliance (ASME BPVC, B31, AWS), tolerance stacks across sheets and interfaces, and drawing integrity: datum errors, contradictions, missing callouts. Huntscreens, a product directory, lists the same four critical checks and notes the evidence‑chain tracking that makes every review auditable and repeatable.

Performance claims are specific. The company says review time drops from days to minutes. Y Combinator's company page puts it at "up to 20× faster" for bringing products to market. "While the current review takes days, Hera takes minutes… and gets better by the drawing." Every review feeds a system of record for drawing errors, which the team says helps build better models for engineering processes over time.

Early customers span data‑center ancillary equipment, renewable‑energy systems, robotics, heat exchangers, power plants, and pressure vessels, all engineer‑to‑order manufacturers where a single missed tolerance callout can scrap a $40,000 part and blow a shipment schedule. The founders' ask is direct: "If you run or work at a factory… we want to visit you." They'll show up in person with a customized Hera Swiss Army knife.

Siemens and the Other HERA

The Siemens collaboration in the research is not a bilateral agreement with Hera AI the startup. It is a multi‑year, multi‑partner Clean Aviation initiative also called HERA (Hybrid‑Electric Regional Architecture) in which Siemens Digital Industries Software works alongside Leonardo and other consortium members to define the digital backbone of a hybrid‑electric regional aircraft. The name overlap is exact, and the aerospace pull is real, but the counterparty is different from the one the hiring surge might suggest. Any hiring surge at Hera AI the company feeds a drawing‑review product; the Siemens partnership documented here advances a virtual verification framework for an entirely separate HERA project.

Siemens Digital Industries Software's role in the Clean Aviation HERA project is specific and technical. The project description says SISW will "contribute to the development of digitally integrated hybrid electric aircraft virtual verification framework, simulation technologies improvements, deployment of MDO workflows supporting design studies from components to aircraft level and specification of future testing for clean aviation phase 2." In practice, that means extending the Simcenter portfolio to model hybrid‑electric propulsion systems end‑to‑end (thermal, electrical, and aerodynamic) while keeping data exchange open through standard formats and interfaces. The work is funded under Clean Aviation Phase 2, targeting high Technology Readiness Levels and a demonstration strategy for a hybrid‑electric regional aircraft entering service around 2035.

On the manufacturing side, Siemens has already deployed its Closed Loop Manufacturing (CLM) solution within the HERWINGT wing‑demonstrator work package feeding the broader HERA effort. CLM links Teamcenter PLM with Opcenter MES so the Outer Wing Box's Bill of Process — every operation, tool, and inspection — stays synchronized from design through production. A September 2025 LinkedIn post from Siemens' Changik Jeon notes that this integration "optimizes the entire manufacturing process, from precise electronic work instructions to detailed key bill of process management and even the critical shimming process for aircraft skin panels." The HERWINGT project site calls the result "a robust foundation for advanced manufacturing planning and execution" that reduces the manual rework plaguing new‑product introduction in aerospace.

Simulation‑driven design is already showing measurable gains. A March 2026 post by Siemens' Will Kruspe reports the consortium's digital‑twin approach has "cut cruise fuel burn by 5% and boosted efficiency, all while staying within 3% of real‑world performance data." Those numbers come from MDO (Multidisciplinary Design Optimization) loops coupling Simcenter's system simulation with high‑fidelity CFD and FEA — exactly the workflow Siemens committed to maturing under the HERA project. The same post frames the work as "a bold step toward net‑zero aviation," reflecting the Clean Aviation program's climate targets.

Beyond the immediate consortium, Siemens is also a core partner in ODE4HERA (the Open Digital Environment for Hybrid‑Electric Regional Architectures), which Clean Aviation describes as "the Digital Backbone of Tomorrow's Aircraft Development." The goal is a transferable, open digital platform letting any OEM or supplier plug in tools for requirements management, verification, and certification evidence generation. Siemens' contribution mirrors its Simcenter openness pledge: provide the simulation and data‑management layer so the ecosystem converges on a single source of truth rather than exchanging STEP files and PDFs.

For Hera AI the company, the aerospace pull is indirect but tangible. The same GD&T, tolerance‑stack, and ASME/ISO compliance checks its drawing‑review tool automates are the bread‑and‑butter of the CLM and MDO workflows Siemens is hardening for hybrid‑electric programs. If a Tier‑1 supplier adopts Hera AI to clear drawing bottlenecks before releasing parts into a Siemens‑managed CLM pipeline, the handoff becomes cleaner, but that integration has not been announced, and the research records no commercial agreement between the two HERAs. The hiring surge in San Francisco builds a product for engineer‑to‑order job shops; the Siemens partnership builds a certification‑grade digital thread for a 2035 aircraft. They share a vocabulary, not a contract.

Incumbents Respond: Autodesk and Rockwell

Autodesk's answer to the drawing‑review automation wave is Neural CAD, a foundation‑model family unveiled at Autodesk University 2025 and rolling into Fusion and Forma. The pitch: "completely reimagine the traditional software engines that create CAD geometry" and "automate 80 to 90% of what you [designers] typically do," according to Autodesk's announcement. Unlike large language models that describe geometry or image models that render it, Neural CAD is trained on professional CAD data so it can reason about boundary‑representation topology, parametric history, and the design intent embedded in constraints — capabilities Autodesk Research has spent more than 15 years building, dating back to the AI Lab's founding in 2018 and the earlier Project Bernini work. The first shipping features (AutoConstrain in Fusion and the Forma Building Layout Explorer) already generate editable B‑rep geometry and, for many Euclidean parts, a full Fusion feature tree with timeline. Autodesk positions Neural CAD as complementary to parametric modeling, not a replacement, and says future interfaces will blend prompts, sketches, reference documents, images, and voice rather than rely on text alone. The company emphasizes trust tooling: internal benchmarking, anti‑parroting checks to prevent reproducing customer designs, transparency documentation, and customer controls over data usage and model opt‑outs. CEO Andrew Anagnost frames the broader vision as "project intelligence" — AI that lowers barriers to specialist expertise so architects and engineers influence more of the project lifecycle.

Running in parallel, Autodesk Assistant (an agentic AI partner) is being embedded across Fusion, Vault, and other products. In Fusion it automates sketch constraints and manufacturing toolpaths while handling team collaboration invites; the company says it "understands your models, your project context, and the task in front of you." AI drawing automation in Fusion is also marketed as a "significant leap forward" for technical drawing creation across industries. These moves predate Hera's public launch but land in the same quarter Hera's drawing‑review tool gains aerospace traction, giving customers two very different automation paths: a specialized QA layer that checks human‑authored drawings against GD&T and tolerance stacks, versus a generative stack that aims to produce the geometry and its documentation together.

Rockwell Automation pivoted hard toward "Agentic AI" in August 2026 alongside a $2 billion U.S. manufacturing infrastructure commitment. The centerpiece is TechConnectIQ Support, a modernization of the 20‑year‑old TechConnect remote‑support service. The new MyRockwellAutomation TechConnect hub surfaces real‑time installed‑base data and digital insights so manufacturers can spot support trends and make faster decisions. AI‑enabled tools (federated search, conversational AI, augmented‑reality guidance, and instructional video retrieval) sit on top of that data layer to cut resolution time and address the skills gap on the plant floor. Rockwell's framing is less about design automation and more about keeping existing automation running: the company says the digitally enabled experience "helps manufacturers address skills gaps, simplify updates and resolve issues faster." The $2 billion capital plan signals a bet that software‑led industrial intelligence, not hardware volume, will drive the next growth phase.

Neither company's announcements name Hera, and the research establishes no direct causal link between Hera's hiring surge or Siemens partnership and these product launches. The timing, however, is notable: Autodesk's Neural CAD reveal at AU 2025 and Rockwell's Agentic AI pivot in August 2026 bracket Hera's public emergence from the batch and its Siemens Digital Industries Software collaboration. An incumbent response layer is forming — generative geometry at the design stage from Autodesk, agentic operations support at the plant stage from Rockwell — while Hera targets the narrow, high‑stakes gap between them: automated verification of 2D drawings before metal is cut.

Policy as Recruiting Tool

The federal tax code has quietly become a recruiting tool for AI startups building in manufacturing. Hera's decision to concentrate its industrial software engineering hires in San Francisco sits against incentives that reward domestic R&D spend, offset payroll costs for early‑stage companies, and subsidize capital equipment, levers that make a high‑cost talent hub more defensible.

The Research and Development Tax Credit, permanent since the 2015 PATH Act, lets startups with under $5 million in revenue and fewer than five years of operations apply credits against the employer portion of Social Security and Medicare payroll taxes. The 2023 expansion doubled that offset to $500,000 annually. Most AI startups see a net benefit equal to 6–10 percent of qualified R&D spend, according to Burkland Associates' 2025 guide. For a company paying San Francisco engineering salaries (where Zero G Talent's board data shows machine‑learning roles banded at $212,000–$318,000), that offset can cover one to two full‑time hires per year.

Section 174 changed the calculus further. Since 2022, R&D expenses must be amortized over five years instead of deducted immediately, raising effective tax rates on engineering payroll. The credit became "more valuable than ever to offset upfront costs," as HRlogics noted in its 2025 manufacturing tax guide.

The CHIPS and Science Act added a parallel track. The Advanced Manufacturing Investment Credit offers 25 percent of qualified investment for semiconductor facilities, rising to 35 percent under the One Big Beautiful Bill Act for property placed in service after 2025. While Hera builds software, not fabs, its aerospace customers (Siemens among them) sit downstream of the $500 billion in private semiconductor commitments announced since 2022, which Deloitte projects will triple domestic capacity by 2032 and create half a million jobs. The policy signal matters: federal money follows domestic manufacturing intelligence, and Hera's drawing‑review tool plugs directly into that stack.

Bonus depreciation remains in phase‑down. One hundred percent expensing applied to qualified property acquired after September 2017 and placed in service before January 2023 (2024 for longer‑lived assets), dropping 20 percentage points per year thereafter. For a startup buying GPU clusters to train drawing‑review models, the declining allowance raises the effective cost of compute infrastructure each year, another reason to front‑load hiring and capital spend.

State programs layer on. The Work Opportunity Tax Credit offers up to $9,600 per eligible hire. Georgia's Job Tax Credit provides up to $4,000 per job annually. New York's Youth Jobs Program pays up to $7,500 per young worker. California runs its own R&D credit parallel to the federal one, though it does not allow payroll‑tax offsets. San Francisco's generative AI sector attracted significant venture capital in 2024 but faces a talent shortage "that money alone cannot solve." The combined federal‑state offset stack narrows the gap between Bay Area compensation and cheaper metros.

The Joint Committee on Taxation's 2021 hearing document on domestic manufacturing incentives (still the baseline for current law) frames these provisions as a system: expensing, credits, and depreciation working together to lower the marginal cost of U.S.‑based innovation. Hera's founders, Meera Patel and Noelle So, come from manufacturing backgrounds; they know the tax code the way shop‑floor veterans know tolerance stacks. This surge is not just a talent play. It is a policy‑arbitrage play, and the incentives are written to reward exactly the kind of engineer‑to‑order manufacturing intelligence Hera sells.

Early Impact: What the Numbers Show

Hera's core claim is straightforward: drawing review that once consumed days now finishes in minutes. The company's own materials repeat the benchmark — days versus minutes — and specify four automated checks: GD&T conformance, code compliance, tolerance stacks, and drawing integrity. When a missed callout escapes manual review, "it can cost weeks of rework on the shop floor." That asymmetry — minutes of compute versus weeks of rework — is the economic argument Hera sells to engineer‑to‑order manufacturers.

Concrete, named‑adopter ROI figures for Hera itself are thin. The company launched publicly in 2026 in the batch. Early customers have not been disclosed with quantified case studies. What exists instead are category‑level proof points from parallel manufacturing‑AI deployments: data showing what becomes possible when AI moves from pilot to production line.

Cadence, whose agentic platform automates PCB and advanced‑packaging design, reported in July 2026 that early adopters including Nvidia, TSMC, Schneider Electric, Socionext, and Forvia Hella saw up to 15× productivity gains and twice‑as‑fast time to market. Forvia Hella reduced a component‑placement task from four days to about four minutes. TSMC, deploying Nvidia AI across its semiconductor fabs, documented 20–50% cost improvement for computational lithography via cuLitho and 50× faster chemistry simulations via cuEST as of June 2026. In logistics, Siemens and Humanoid Ltd. recorded 60 container moves per hour with greater than 90% pick success in a live Erlangen factory trial reported in August 2026. Agility Robotics' Digit humanoids have moved 100,000 totes at a GXO facility and secured $300 million in multi‑year contract orders across Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada.

These numbers are not Hera's. But they establish the performance envelope Hera targets: the same shift from days to minutes, the same elimination of rework loops, the same auditable traceability that regulated aerospace and defense buyers require. Hera's Siemens Digital Industries Software collaboration (detailed in the previous section) is the channel through which those aerospace verification workflows will get their first at‑scale test. The partnership contributes to the framework and MDO workflows from component to aircraft level, per the Clean Aviation Phase 2 scope.

The ROI logic for early Hera adopters will likely follow a three‑step pattern visible in the Cadence and TSMC data: first, review‑cycle compression (days to minutes); second, escape‑rate reduction (fewer missed callouts reaching the floor); third, capacity unlock (senior engineers redeployed from red‑pen checks to higher‑value work). The "Live in a week" deployment claim suggests the integration friction that Caterpillar's Mineart flagged ("the hard part about autonomy and about physical AI is incorporating that technology into the customer jobsite and into the workflows") has been engineered down for the drawing‑review use case.

What remains unproven is whether Hera's specific GD&T and tolerance‑stack math engine delivers the same magnitude of gain in low‑volume, high‑mix aerospace job shops that Cadence sees in high‑volume PCB flows. The Siemens partnership is the designated proving ground. Until named aerospace customers publish cycle‑time and scrap‑rate deltas, the ROI case rests on the category precedent and Hera's architectural claims: It reads and checks the answers, the engineer still signs. The numbers will come from the shop floor — not the pitch deck.

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