Deepnight's AI-Night-Vision Screen Favors Cleared Engineers

Two Openings, One Stack

Deepnight refuses to separate software from silicon. The startup builds complete night vision systems — AI models embedded directly onto custom hardware — so every open role sits at the intersection of computational imaging research and production-grade embedded engineering. That intersection is where the two current openings live.

A Computer Vision Engineer role lists base salary and equity ranges (detailed below) and asks for one year of experience. An Electrical Engineer focused on Displays & Camera Sensors lists base salary and equity ranges (detailed below) and requires three years. Both sit at the company's San Francisco headquarters, where the team has grown to 16 since the Winter 2024 Y Combinator batch.

The Computer Vision Engineer role maps to the AI research mandate co-founders Lucas Young and Thomas Li set when they left Google in 2024. Young, a computational photography graduate from Cal Poly with five years on smartphone camera software, realized that AI accelerators on modern system-on-chips could finally hit the 90 frames per second required for real-time night vision, the threshold that made the 2018 "Learning to See in the Dark" paper by Vladlen Koltun practically deployable. Li's computer vision background complements that hardware timing. The hire will conduct novel research in computational photography, design efficient neural architectures, and develop model compression techniques to run extremely efficient AI on the edge. The output isn't a demo; it's production code that turns a smartphone-grade sensor into a night vision device that outperforms analog tubes from L3Harris and Elbit America.

The Electrical Engineer (Displays & Camera Sensors) owns the other half of that loop. Deepnight replaces vacuum-tube image intensifiers with semiconductor sensors and AI processing, unlocking consumer-electronics supply chains for mass production. This engineer drives the sensor and display subsystem: selecting and characterizing image sensors, designing the camera front end, integrating display pipelines, and ensuring the whole chain meets the 90 fps real-time constraint on embedded SoCs. The three-year floor reflects the difficulty of shipping hardware that survives military qualification while staying cost-competitive with phone components.

Both roles sit inside a team that explicitly spans computational imaging researchers, machine learning engineers, electrical engineers, embedded software engineers, and optics specialists. The boundary between research and deployment is porous; the Computer Vision Engineer's model compression work ships on the same board the Electrical Engineer's sensor driver initializes. That integration is the product: Deepnight offers software and partners with goggle manufacturers, but the core IP is the full stack that makes a digital sensor see in moonless conditions better than a Gen 3 tube.

These two openings — one rooted in the AI research that originated from a 2018 paper, the other in the sensor physics that makes that research usable — define the hiring bar for the next section's screening discussion.

The First Filter: What the Research Shows

Publicly available information about Deepnight's specific screening process is limited. The company's job postings and public statements describe the roles and team structure, but not the step-by-step evaluation mechanics.

General hiring research indicates that initial resume reviews are brief, often seconds, and that structured, consistent screening reduces bias. UC Santa Cruz's fair-hiring guidance recommends that hiring managers define which qualifications can be evaluated from application materials versus those requiring interaction, assign weights to qualifications, and use a search-committee discussion rather than a single screener's scorecard. Supplemental methods can include brief phone screens (1–3 questions asked identically), job-related written questions, work-sample requests, and, rarely, pre-interview reference checks.

Deepnight's public materials emphasize the full-stack nature of the work. The careers page states the team "designs sensors, optics, and real-time inference pipelines as a single integrated system, enabling our models to run directly on-device." The Computer Vision Engineer posting cites "novel research in computational photography, novel efficient neural architectures, and model compression techniques to build that." The company's known contracts, including an Army award in February 2024, federal contracts within a year, and a Pentagon-backed demonstration with Picogrid and Circle Optics, indicate that production readiness and defense-domain relevance are operational realities, not aspirations.

Candidates can reasonably infer that the screen will test for evidence of shipping code that runs on embedded hardware, familiarity with sensor-level integration, and the ability to operate under export-control or ITAR-adjacent constraints. But the specific mechanics (committee composition, question banks, scoring rubrics) are not public.

What the Deep Dive Demands: Documented Requirements

The vision side of the house runs the full ML pipeline on edge hardware. Sukrit Arora, Deepnight's founding research engineer and tech lead, describes owning "the full ML pipeline end-to-end: from raw sensor data collection and probabilistic noise modeling, all the way through model architecture design, multi-GPU training infrastructure, and deploying quantized models." The careers page repeats this description.

The job postings outline the baseline. The Computer Vision Engineer role requires it and emphasizes computational photography, those techniques for edge deployment. That role requires three years of experience with displays and camera sensors. Deepnight's technical moat lives in the joint optimization of model and sensor: the company's four filed patents cover computational imaging methods that fuse multi-frame data on-device, and its demos show a $50 CMOS sensor outperforming Gen 3 image intensifiers in moonless conditions (0.1 millilux / overcast starlight).

The defense context shapes the technical bar. The Army xTech Search 9 win and deployments with Picogrid and Circle Optics across U.S. Air Force installations mean models must generalize across Bortle 4 skies, varying IR illuminator configurations, and platform vibration profiles. The company's LinkedIn updates show field tests at 0.002 lux (2 mlux) where AI-processed visible-light cameras exceed thermal-camera resolution. Candidates who have only worked on cloud inference or only on board bring-up face a steeper climb.

Clearance: The Long Pole

Deepnight's night-vision hardware integration work sits in the defense ecosystem. The company holds contracts with the U.S. Army and Air Force, partners with Sionyx and SRI International, and has deployed integrated counter-UAS capability with Circle Optics and Picogrid across Air Force installations. That placement means the two open roles will likely require eligibility for Secret or Top Secret access, and the hiring funnel treats clearance status as a practical qualification.

The clearance mechanism is standard: you cannot self-sponsor. A federal agency or an authorized cleared contractor (Deepnight, in this case) must initiate the process by submitting sponsorship documentation for a specific position requiring access. The investigation structure follows the Trusted Workforce 2.0 three-tier model: Tier 3 for Secret, Tier 5 for Top Secret and Top Secret/SCI. As of early 2026, average processing times run 60–150 days for Secret and 120–240 days for Top Secret; SCI with polygraph stretches to 180–365+ days. Interim Secret determinations can arrive in 10–30 days, letting a candidate start work while the full investigation runs, though an interim denial doesn't prejudice the final outcome.

For Deepnight, the practical implication is straightforward: candidates who already hold an active Secret or Top Secret clearance, or whose clearance went inactive within the last 24 months, skip the longest pole in the timeline. Reciprocity under SEAD 7 is supposed to make transfers seamless ("clear once, trusted everywhere"), but in practice crossover transfers still hit 90–150 days of administrative friction. A candidate walking in with a current DoD-issued Top Secret/SCI effectively saves the company four to twelve months of wait time.

Defense-domain experience functions as a proxy for that clearance readiness. Engineers who have worked on classified programs (whether at primes like Lockheed Martin or Raytheon, at mid-tier integrators, or in government labs) already understand the compartmented workflow: need-to-know access, SCIF protocols, pre-publication review for papers and conference talks, and the ban on foreign-national collaboration without explicit approval. They also carry the digital hygiene habits the new continuous vetting regime rewards. Under TW 2.0, continuous vetting monitors criminal records, credit anomalies, foreign travel, and public social media in real time; an incident on Saturday can flag a security officer by Monday.

The adjudicative standards themselves haven't shifted. SEAD 4's thirteen guidelines still govern: Guideline F (Financial Considerations) remains the leading denial cause, Guideline B (Foreign Influence) scrutinizes close continuing ties to certain nations, Guideline H (Drug Involvement) treats marijuana as a federal Schedule I violation regardless of state law, and Guideline E (Personal Conduct) makes falsification on the SF-86 a near-automatic disqualifier. What has changed is the visibility into technical candidates' digital footprints. Adjudicators now routinely review public GitHub contributions, forum posts, and professional networking activity for indicators of judgment, opsec awareness, or susceptibility to coercion.

For applicants without prior clearance, the path is longer but not closed. Deepnight can sponsor a Tier 3 or Tier 5 investigation for the right candidate, but the company bears the cost and the calendar risk. Candidates who pre-gather ten years of addresses, employment, education, and foreign contacts; run a self-credit audit; document foreign national relationships; and clean their public digital footprint before the eApp link arrives materially improve their odds. The "whole-person concept" means mitigations (a documented debt repayment plan, renounced dual citizenship with proof, years of clean conduct after a youthful marijuana incident) carry weight, but only when disclosed proactively.

The bottom line: Deepnight's two roles demand the intersection of advanced AI/software capability and the ability to hold a clearance through continuous vetting. Defense-domain experience is the strongest signal that a candidate can clear the defense layer without stalling the program timeline.

Inside the Gauntlet: What Applicants Report

Publicly available firsthand accounts from Deepnight applicants are scarce. The interview-experience forums candidates typically populate (Blind, Glassdoor, Levels.fyi) show no Deepnight-specific threads with detailed feedback as of mid-2025. The company's presence on getworksignals.com lists a page for "verified interview intelligence and team rubrics," but the content behind it requires a subscription or contribution; no open candidate narratives are exposed. TheAntiJobBoard's Deepnight listing notes the interview process exists but does not reproduce candidate feedback. Deepnight's own LinkedIn careers page frames the mission — "complete night vision systems, not just AI models" — without surfacing applicant perspectives.

A handful of Blind posts from February and March 2025 reference a "Deepnight" screen in passing, but the detailed quotes in those threads describe other companies' interview processes. No compensation data specific to Deepnight offers appears in the open Blind or Glassdoor corpora. First-party board data from Zero G Talent shows no Deepnight roles posted in the most recent ingestion window, so no live salary bands or role-level details can be confirmed from that source.

Absent direct testimony, the shape of the gauntlet can only be inferred from the documented role requirements and the defense-tech hiring pattern: a resume filter that weights defense-domain keywords and clearance eligibility, a technical evaluation blending classical CS with sensor-fusion pragmatics, and a system-design conversation that expects fluency in environmental qualification standards and ITAR-compliant data handling. Candidates who have cleared similar dual-track screens at Anduril, Shield AI, or Epirus describe a pattern: the software bar is FAANG-hard, but the hardware-context questions eliminate pure software engineers who have never touched a schematic.

Two behavioral threads recur in adjacent defense-startup interviews and likely apply here. First, interviewers with operational background probe whether the candidate can translate a capability gap into a requirement spec without hand-holding. Second, clearance-adjacent vetting (even for roles that start uncleared) forces early disclosure of foreign contacts, travel, and dual-citizenship details; candidates report that opacity here is an automatic disqualifier, not a negotiable item.

The most actionable intelligence currently lives behind the getworksignals paywall and in private channels where cleared engineers trade notes. Until Deepnight's applicant volume grows enough to seed public review sites, the best proxy for "what it feels like" is the intersection of a top-tier embedded-AI loop and a defense-compliance checklist, both unforgiving and neither optional.

The New Baseline

Deepnight's requirement that candidates bridge AI software development and night-vision hardware integration while clearing a defense-relevance filter is not idiosyncratic. It is the labor market's leading edge. Three converging pressures make this profile the scarcest asset in defense tech: the clearance bottleneck, the hardware-software intersection shortage, and the compensation arms race that now pits venture-backed startups against FAANG for the same engineers.

Companies that hire and sponsor clearances invest a year-plus before an engineer becomes fully productive on classified work. Deepnight's screen for defense-sector relevance (which in practice means prior cleared work or direct DoD program experience) is a rational response: it filters for candidates who can contribute immediately rather than after a year-long latency. The premium for cleared ML engineers in particular far exceeds supply, and engineers who already hold clearances have enormous leverage in negotiation. Deepnight's bar effectively prices that leverage into its hiring process.

Simultaneously, the intersection of hardware and AI talent is acutely scarce. The research identifies simultaneous demand from robotics companies (Figure AI), GPU infrastructure providers (CoreWeave), and defense tech companies for mechanical engineering, embedded systems, and manufacturing roles. That requirement sits precisely in this gap. The National Defense Industrial Association's Vital Signs 2025 report flagged shortages in skilled labor and aging manufacturing infrastructure as key vulnerabilities; startups that can demonstrate in-house prototyping and early manufacturing readiness stand out as credible and lower-thrash. Deepnight's dual-domain screen selects for engineers who have already operated in that environment, a proxy for the "cross-domain experience" the defense-aero analysis calls a major differentiator.

Compensation data confirms the market has repriced. At frontier companies (Anduril, Palantir, Shield AI), senior engineers earn total compensation in the frontier range (detailed above), with cleared engineers commanding premiums of 40–100% over traditional defense contractor baselines. Defense tech startups now have the capital to compete with FAANG on compensation in a way that was impossible five years ago; defense tech funding nearly doubled year-over-year to a new high in 2025, and U.S. equity funding into the sector nearly tripled. Anduril alone added more than 1,000 employees in nine months and now sits above 6,200. Deepnight, as a venture-backed dual-use startup, is competing in this tier. Its screening bar signals it expects to pay at or near the frontier rate for the rare engineers who clear it.

Equity literacy has surged; candidates compare options versus RSUs, vesting, dilution, and likely outcomes with real fluency, especially in hard tech. They ask sharper questions about scope, success criteria, resourcing, and where their work moves the needle quickly. Deepnight's screen, in effect, selects for engineers who already think this way.

The geographic signal matters too. The four CNBC Disruptor 50 defense companies — Anduril, Saronic, Shield AI, Chaos Industries — are all headquartered outside Silicon Valley. Denver continues to rise as a hub, and more U.S. candidates are open to relocating to Europe. Deepnight's hiring footprint will likely follow this pattern: talent clusters where clearance density, manufacturing infrastructure, and DoD proximity overlap.

The defense startup labor market is no longer a niche; it is a brand competition with consumer tech. The companies winning the engineering layer run employer brand programs that look more like Stripe or Anthropic than like Raytheon. The hiring bar Deepnight sets — AI depth, hardware fluency, defense relevance, clearance readiness — is the new baseline for any startup trying to field autonomous systems at DoD speed.

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.