Tiburon Labs - Confidential

Premise

America is entering an era of generational infrastructure buildout: power, broadband, data centers, advanced manufacturing, and other critical infrastructure all need to scale at once. Capital is available, demand is clear, and the national imperative is increasingly critical. But these sectors share a common bottleneck: 3M+ workers are needed in the US by 2035, but we cannot produce trusted technical execution capacity fast enough.

The limiting factor is not simply labor supply. It is the slow, supervision-heavy process of turning available workers into people who can reliably execute real jobs in the field - safely, correctly, and with documentation that passes acceptance. The problem is compounded by the fact that 30% of the current workforce is retiring in the next decade - taking tribal knowledge that today has no scalable way to transfer.

This problem has existed for decades. What has changed is that multimodal AI is approaching a threshold where real-time visual reasoning about physical work becomes feasible. Vision-language models can now perform zero-shot segmentation at near-production quality, learn new tasks from a few examples without retraining, and run lightweight decoders on-device while offloading heavy reasoning to the cloud. For the first time, it is becoming possible to build systems that observe, understand, and guide physical work in real time.

Tiburon Labs exists to close this gap. We believe AI should not replace human reasoning in these environments, but augment it: to help workers reason better in context, to help experts supervise more effectively, and to make real work a source of continuous learning and proof.

If AI can turn raw labor supply into trusted execution capacity, it becomes a foundational technology for building industrial capacity at the scale we critically need.

Why This Matters

At a moment when much of the public conversation about AI is centered on job displacement, we are interested in the opposite problem: how AI can expand access to the workforce.

Skilled trades are one of the most reliable paths to the middle class in America. It is how someone goes from earning $30,000 a year to earning $70,000, from renting to owning, from surviving to building something for their children.

But opaque career pathways, long apprenticeships, and limited training capacity keep this path too narrow and too slow.

We want to open it up for millions more people.

The Opportunity

The labor problem in these sectors is often misunderstood. The issue is not that every task requires a deeply experienced expert. The issue is that most workflows combine a large amount of structured, repeatable execution with a much smaller amount of scarce supervision, signoff, troubleshooting, and exception handling. That creates three structural bottlenecks:

  1. Upskilling is too slow. Existing technicians take too long to ramp into new procedures, new equipment, and higher-level work because knowledge is fragmented across manuals, tribal know-how, and local supervision.
  2. Entry-level workforce expansion does not scale. Even if organizations could hire more L1 workers, supervision capacity becomes the constraint. Foremen and senior technicians do not have the bandwidth to oversee large numbers of novices while still completing their own work.
  3. Coordination is too brittle. SOPs change, procedures evolve, customer requirements shift, and organizations struggle to propagate those updates quickly and consistently across distributed teams.

These are not separate problems. They are symptoms of the same missing layer: there is no system that can understand work as it happens and structure execution across the workforce.

From The Field

We have spent months talking to people who live with this problem.

A staffing executive responsible for AT&T transmission line projects described the cycle: AT&T requests labor. They assess the job and commit to 150 technicians. They spend 6 to 12 months finding and training people for that specific project. Then AT&T announces the project was delayed or the region changed. Massive attrition follows. Only 20% of people stay because they need to find other work and cannot wait.

At TSMC, process engineers manage technicians who must adhere to strict SOPs. Today, technicians often look back and forth between the equipment and PowerPoint slides describing the procedure. When something is unclear, they flag a process engineer who reviews the issue later. The technician waits. In a manufacturing environment where minutes of delay matter, execution is still coordinated through PowerPoint and follow-ups on Microsoft Teams.

At Oracle data centers, tasks are assigned on Jira. No photos. Reactive verification only after problems emerge. At Sabre Industries, verification still happens in-person by foremen who don't have the bandwidth.

What We're Building

Tiburon is an AI-guided execution, pedagogical, and verification system for skilled technical work.

A worker carries or wears a device while performing a task. The system observes the work session, maintains a live understanding of the job state, guides the worker step by step, and captures evidence of what was completed. If something is uncertain, unsafe, or outside policy, the system asks for more evidence, slows down, or escalates to a senior technician.

Underneath this is what we call a Procedural Intelligence Layer - a live belief about the state of a physical job:

The goal is not to be merely instructional, but pedagogical: the system helps workers understand the work, build judgment, and move up the skill ladder through execution itself. Every work session becomes both productive output and a learning episode.

Built this way, the product is useful to the workforce from day one - shortening ramp time for new technicians, extending the leverage of senior ones, and giving organizations a verifiable record of every job. The data it generates, captured as a byproduct of real work rather than solicited as demonstrations, is the foundation for the models that will eventually run alongside humans in real time.

That data is uniquely valuable because trades work is governed by information that does not appear in the visual stream. A camera captures geometry and motion, but does not capture what makes the motion correct. That hidden information shows up in two layers:

The Data Layer

Every escalation produces a structured trace of how the work was actually performed. The trace pairs three things that don't currently exist together at scale:

Most of what a senior technician knows transfers across employers: the physics, the failure modes, the reasoning patterns. What does not transfer is the company-specific layer of SOPs, equipment manifests, and acceptance criteria. We surface that layer as context with each escalation, and over time the model absorbs it. Experts are needed only for genuinely novel cases.

Those novel cases are the edge cases - rare failure modes, unfamiliar equipment, unusual field conditions - and they consume a disproportionate share of diagnostic time and rework. No single organization sees enough of them to learn from, but the network does. The corpus compounds with every contractor, equipment model, and field condition added.

How We Start

To build intelligence that understands physical work, you first need data about how physical work is actually performed. That data does not exist anywhere in structured form. We are creating this data engine as a byproduct of a product technicians genuinely want to use.

This is the entry point. The product evolves in four stages, each building on the data and capabilities established by the last.

Act 1: Expert-Guided Execution + Data Capture. A novice technician wears a head-mounted camera (glasses, GoPro, or similar egocentric device) while performing a task. A remote expert - an experienced or semi-retired technician - observes the live feed, guides the worker through each step, and explains the reasoning behind every judgment call. Every session is captured: synchronized egocentric video and expert audio narration, structured into action-intent annotations. This is the data engine - each deployment captures tacit knowledge in structured form for the first time.

Act 2: AI Trained on Expert Knowledge. With enough expert sessions logged, the system no longer relies on human guidance alone. It has learned from thousands of real resolutions - how experts actually diagnose problems, what the manuals miss, what only experience teaches. The AI begins handling routine guidance directly through the same wearable form factor. Escalations to human experts get fewer. The flywheel accelerates.

Act 3: Autonomous Real-Time Guidance. The system now reasons about live physical work in real time, hands-free - powered by models trained on everything captured in Acts 1 and 2. AR overlays and contextual prompts guide the technician through procedures, verify completion, and escalate only when genuine uncertainty or novel situations arise.

Act 4: Execution Infrastructure. Tiburon becomes the system through which firms recruit, onboard, qualify, supervise, coordinate, and verify technical labor - the operating layer for how physical work is executed and learned across the workforce. [see: The Bigger Picture]

Where We Start

Our view is that the best initial wedge is a trade where three things are true at once: the labor shortage is severe, the workflows are structured enough for AI guidance, and the buyer base is concentrated enough that success compounds commercially.

Fiber is the strongest starting point. It has some of the lowest credential gates of any skilled trade - most roles require only a high school diploma and a driver's license, with most training done on the job. The addressable labor pool is large and the barriers to entry are low, but the barriers to productivity are not. Industry research shows it takes three to six months for an inexperienced hire to produce work of sufficient quality. The Fiber Broadband Association projects 205,000 new fiber technicians will be needed in the next five years, with no credible pipeline to produce them. Over $42B in federal broadband funding is actively flowing, with subgrantees legally required to deliver service within four years or face clawbacks. Verizon and AT&T have committed a combined $25B+ in private investment on top of that. The urgency is structural and the capital is already committed.

Through our partnership with the Fiber Broadband Association and their 830 partner organizations, and our deployment with DPR on active data center construction, we project collecting several hundred hours of expert-annotated egocentric video in the first 12 months, across 5-10 core fiber task types. Each hour produces structured action-intent annotations - not just "splice fiber" but the full causal chain: what the technician checked first, why they chose a particular technique, what they verified after completion, and the verbal reasoning throughout.

How We Scale

Each trade we master applies horizontally across every industry it serves. A fiber model doesn't just work in telecom - it works in data centers, factories, and military bases. Each new trade adds a new domain of structured procedural knowledge, and each deployment adds hours to the corpus.

Data centers, energy infrastructure, manufacturing, oil and gas, healthcare, defense, and commercial construction all rely on overlapping combinations of the same trades. We don't enter industries one by one. We enter them by mastering the trades they all share.

Fiber → Electrical → HVAC/Mechanical → Pipefitting → Controls → Industrial Maintenance.

The Bigger Picture

At its core, Tiburon is an operating system for how technical work is executed, verified, and learned in the physical world.

Labor is often 20-40% of the budget on large infrastructure and industrial projects. Today, that labor component is difficult to manage with precision because execution quality, supervision burden, skill distribution, and rework risk are all hard to observe directly. If Tiburon can make work machine-guided and machine-verified, it creates a much richer substrate for understanding and managing labor productivity itself.

Over time, Tiburon can become the system through which firms recruit, onboard, qualify, supervise, coordinate, and verify technical labor across distributed teams and complex projects. That alone is a very large software business, and the structured execution data it generates has value far beyond any single deployment.

Team

Shree Reddy - Stanford CS M.S. + B.S. Worked on compute market design at Compute Exchange, studying on the ground how workforce and operational bottlenecks constrain infrastructure buildout. ML researcher at Radical Numerics (founded by the Evo authors from Arc Institute) and Stanford's Scaling Intelligence Lab with Prof. Azalia Mirhoseini, working on recursively self-improving AI systems; previously infrastructure engineering at Glean.

Riya Karumanchi - Stanford CS M.S. + B.S. Featured on TIME in 7 Young Inventors Who See a Better Way. Worked on industrial capacity and advanced manufacturing research for Prof. Drew Endy, materials science research in spectroscopy at the Dionne Group, and on machine learning systems for real-world sensing and instrumentation at Pumpkinseed Technologies. Previously worked on cybersecurity infrastructure strategy at McKinsey.

Ira Thawornbut - Stanford CS M.S. + B.S. Siebel Scholar (one of 80 top graduate students worldwide; one of five in Stanford CS). Led a research team building multimodal AI for human physiology funded by Stanford HAI. Research on real-world LLM tutoring at Stanford NLP with Prof. Diyi Yang. Built Thailand's first low-cost ventilator for field hospitals during COVID-19; National Innovation Award, 2020.

FAQ

Is this just training software? No. Tiburon operates during live work. The pedagogical component is embedded in execution itself: the system scaffolds tasks, adapts support, and helps workers build competence through real jobs.

Is this about replacing skilled workers with AI? No. The goal is the opposite: to augment human reasoning, increase expert leverage, and bring more people into productive technical work.

Why is this a large company, rather than a workflow tool? Because the underlying execution data is the missing layer beneath upskilling, supervision, qualification, coordination, compliance, and risk. If Tiburon captures that layer, it can become infrastructure for both workforce markets and capital markets.

Does this require navigating complex licensing or credentialing regimes? A substantial share of technical work is ungated: roughly 40% requires only a high school diploma plus on-the-job training, which allows Tiburon to deliver value without relying on complex licensing frameworks. At the same time, many workflows sit adjacent to licensed roles. Tiburon can also support credentialed technicians by helping them ramp into new procedures, equipment, and job environments more quickly, and by helping organizations transition trained workers into roles where their credentials are required.

Is there willingness to pay for this initial product?

We spoke to one of the five largest mechanical contractors in the country, who have 3,600 field technicians and 170,000+ equipment models in the field. When a technician gets stuck on a job, they call the "Fix Center" - a hacky support line they built and staffed with 8 senior technicians, full time, just to answer those calls. It handles 28,000 calls a year, and every call saves a technician roughly an hour of lost time.

This tells us two things:

  • Organizations will pay to solve this. Comfort is already spending on in-house headcount to manage a problem that software should be solving.
  • Expert guidance drives measurable value at scale. Eight people meaningfully supporting 3,600 technicians is a striking ratio - and a proof point that even zero-AI, reactive intervention shortens ramp time and increases time-to-resolution per field technician.

We also see the retiring workforce differently. 30% of experienced technicians will retire in the next decade. We see this as an opportunity - these are people who built careers on exactly the knowledge that's missing, and many would rather stay connected to the craft than walk away entirely. We are building a network that lets them do that, fielding escalations on their own time and earning income in the process.

If automation is advancing, why focus on training human technicians? Automation will likely affect entry-level physical tasks first, but that does not eliminate the need for technicians. In practice, it increases demand for higher-skill roles that install, integrate, supervise, troubleshoot, and maintain increasingly complex systems.

  • That makes upskilling the workforce more important, not less. As more routine execution becomes automated, the bottleneck shifts to experienced technicians who can oversee systems, handle exceptions, and manage complex infrastructure. Tiburon focuses on accelerating that progression. By guiding structured work, verifying execution, and capturing evidence during real jobs, the system helps workers move from entry-level tasks into more advanced technical responsibilities faster.
  • At the same time, emerging robotics research suggests that egocentric human demonstrations are increasingly useful for training physical AI systems. For example, recent work on large vision-language-action models (e.g., Physical Intelligence's π models) shows that scaled robotic models can leverage egocentric human video during fine-tuning, improving performance on manipulation tasks where robot data is limited.

As physical systems become more automated, the value of technicians who can supervise, diagnose, and manage those systems only increases. Tiburon's goal is to help more workers reach those roles faster.