<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sudeep Fadadu on Medium]]></title>
        <description><![CDATA[Stories by Sudeep Fadadu on Medium]]></description>
        <link>https://medium.com/@sfadadu?source=rss-608c269d0b64------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vpI0calhGrrUT_DAYoC89g.jpeg</url>
            <title>Stories by Sudeep Fadadu on Medium</title>
            <link>https://medium.com/@sfadadu?source=rss-608c269d0b64------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 21:07:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sfadadu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Is career in computer vision engineering a Dead-end ?]]></title>
            <link>https://medium.com/@sfadadu/is-career-in-computer-vision-engineering-a-dead-end-f6c88f572011?source=rss-608c269d0b64------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6c88f572011</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[career-advice]]></category>
            <dc:creator><![CDATA[Sudeep Fadadu]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 00:44:58 GMT</pubDate>
            <atom:updated>2026-04-27T00:44:58.535Z</atom:updated>
            <content:encoded><![CDATA[<p>Until end of last year, despite LLMs on track for becoming world class SWE, I was still fairly confident about job security as a computer vision engineer.</p><p>Then I got a reality check.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1yaFFLYpjx5pdoOb" /><figcaption>Photo by <a href="https://unsplash.com/@redaquamedia?utm_source=medium&amp;utm_medium=referral">Denny Müller</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I participated in a competition at an ICCV workshop. Roughly 30 teams from around the world. Second edition of the competition. The task was to detect and localize road hazards from video. Classic computer vision territory with high complexity.</p><p>I referenced previous year’s results. The top 5 approaches were published at ICCV. Each method was as complex as the next. Temporal filtering, signal processing to downsample key frames, then various ML models on those frames. YOLO, vision transformers, custom architectures, different training strategies, and data processing pipelines. Last year, winning required weeks of engineering, cloud compute, and deep CV expertise.</p><p>I wanted to test a hypothesis. What if I skip all of the complexity. No training. No data augmentation. No cloud GPU setup. Just upload the raw video to a VLM with a simple prompt: extract timestamp, involved actors, and scene description.</p><p>This was supposed to be my baseline, not a final solution. I had ideas for finetuning and filtering to process long video snippets. But the baseline was the top performing method. A single free-tier API call to Gemini Flash 2.5.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*arDoj38_Eb3FtrsEqHOA7g.png" /></figure><p>That result broke something in my head. If a free-tier API call can outperform teams of CV engineers running purpose-built models, is it even worth employing a team of engineers to build a solution to a specific detection problem?</p><p>Do startups even need to hire a computer vision engineer anymore? What is the value in employing a research engineer to build, train, and deploy a model when you can get an API key to a deployed model that requires no fine-tuning? No data gathering. No labeling. Half the jobs that exist to train a dedicated ML model for object detection or scene understanding just evaporated.</p><p>This is not just my experience. I review manuscripts for top CV conferences. The trend is there. Many papers now ask the same question: can VLM X perform CV task Y? The answer is increasingly yes.</p><p>Here is how zero-shot VLM methods are getting close to fully supervised methods on the same dataset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lu904iRI1YRBEc3xWBuz1w.png" /><figcaption><em>Table from ‘Chain of Visual Perception’ (Tang et al.) demonstrating zero-shot multimodal foundation models outperforming fully-supervised architectures on concealed object detection benchmarks.</em></figcaption></figure><p>And in some cases, Zero-shot foundation models consistently outperforming supervised methods trained explicitly on domain data for road anomaly and obstacle detection.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*ElJcT9KdfVXRa9QbN4FXKg.png" /><figcaption><em>Table from Sinhamahapatra et al.’s research proposing the zPROD framework, demonstrating how zero-shot inference using foundation models overtakes dedicated supervised methods in detecting road anomalies and obstacles.</em></figcaption></figure><p>Publication share of zero-shot VLM-based methods has shot up from merely 5% of all object detection papers to almost 25%. Industrial adoption of these approaches is inevitable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*usK2o82QWLADQy7RSSidfQ.png" /><figcaption>Consistent increase in share of published research on Zero short object detection.</figcaption></figure><h4>It Gets Worse.</h4><p>Maybe API-based VLMs are eating up the “train a model for task X” jobs. But surely careers that require building dedicated on-device models are safe? Robotics, self-driving, places where you can not rely on a datacenter across the country with spotty internet?</p><p>I started digging into what is happening in perception for robotics. I got a different story, but same tone.</p><p>The industry is moving from the classic Sense → Plan → Act pipeline to end-to-end models. Pixels in, control commands out. No more hand-crafted perception modules feeding structured data to a planner. Just one model, trained on raw sensor data, outputting motor commands.</p><p>The shift is not subtle.</p><p>Tesla’s FSD v13 deleted over 300,000 lines of C++ code in favor of an end-to-end ML model. That is not a refactor. That is an extinction event for the engineers who wrote and maintained those 300,000 lines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/689/1*NR7yA_oS79Fk0w_96EGxJQ.png" /><figcaption>FSD v12 upgrade notes from my Tesla app.</figcaption></figure><p>New robotics startups like Figure and 1X are not even considering the modular approach. They are starting end-to-end from day one.</p><blockquote><em>From Brett Adcock, Founder and CEO of Figure AI: “… the Figure AI team is running humanoid robots with end-to-end neural nets performing highly complicated and dexterous tasks. these are tasks too complicated to write heuristics in C++ …” on </em><a href="https://x.com/adcock_brett/status/1739719880499425506"><em>X</em></a></blockquote><p>One model vs hundreds of contracts between subsystems. One model to maintain vs 300k lines of C++ spread across teams. Companies care about headcount and spending. An E2E model, once proven, needs fewer engineers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zI09LgMNjaYkSnJbHTDvZg.png" /><figcaption>Head to head comparison (hypothetical) of traditional modular stack with E2E architecture</figcaption></figure><p>And these E2E systems are getting better fast. Simulation from world models and diffusion-based learning policies are generating near-unlimited training data. As data grows, E2E models improve. As they improve, more companies switch. It is a flywheel and it is spinning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fVm4oghVbGoCL1J--u3O4w.png" /></figure><p>Companies that can not go fully E2E are finding a middle ground: multitask learning within perception. One model handling detection, segmentation, depth estimation, and lane detection simultaneously. Fewer models mean fewer GPUs onboard, which means lower headcount cost.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UM6WbKVjKe1ood7Ok01iqw.png" /></figure><p>The skills that got perception engineers hired five years ago (SLAM, feature correspondences, multi-sensor calibration) are being absorbed into the latent space of a model that nobody fully understands. Does this mean the perception engineer as a distinct role might be fading?</p><p>E2E models have limitations for safety critical applications. They lack interpretability. These are valid concerns. But we have heard this exact argument before. In 2016, early self-driving companies were skeptical about using CNNs and kept running SVMs in their perception stack. I do not know a single self driving company still using SVMs today.</p><p>Technology improves. Interpretability tools will catch up. And when they do, the last argument for maintaining factored perception systems may disappear.</p><p><strong>Then what is left?</strong></p><p>If API-based VLMs can replace your task-specific model, and E2E systems can replace your perception stack, what exactly does a computer vision engineer do next?</p><p>When I first started pulling on this thread, the answer felt like probably not a lot except for some legacy companies. The doom is probably real.</p><p>But that is not where the data points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/693/1*wWRe5HEVVcP0z-4V12cmaw.png" /><figcaption>Google Job interest shows a spike in computer vision engineer in last 2 years.</figcaption></figure><h3>So whats going on?</h3><p>Let me zoom out from the models to the market.</p><p>Three years ago, launching a tech startup meant raising a seed round, hiring 10+ engineers, setting up infra, and spending six months before shipping anything. Today a solo founder with Claude Code and a google AI studio account can ship a working product in a weekend.</p><blockquote>(Data backs this up: in recent Y Combinator batches, roughly 75% of funded companies boast extremely lean teams of just one or two founders, tackling scopes that previously required entire engineering departments.)</blockquote><p>What does this have to do with computer vision? A lot probably.</p><p>When starting a company becomes this cheap, more people start companies. When more companies start, the market fragments. When the market fragments into hundreds of niches, each niche needs specialized solutions. Not a general purpose API. A system that works for their specific problem, their specific data, their specific edge cases.</p><p>Think about the kinds of companies popping up that did not exist five years ago:</p><ul><li>Drone inspection companies scanning solar panels for micro-cracks (sub-pixel precision needed)</li><li>AR try-on for jewelry and watches (real-time 3D mesh reconstruction)</li><li>Automated QC for PCB manufacturing (defect detection at 0.1mm)</li><li>Livestock health monitoring from overhead cameras (specialized domain, no public training data)</li></ul><p>Each needs domain expertise, custom data pipelines, and models that run under constraints. Each one is a job posting for someone who understands vision systems.</p><p><strong>The API Honeymoon</strong></p><p>The pattern plays out exactly the same across startups using vision. I watched a logistics startup go through this exact arc recently:</p><p><strong>Month 1</strong>: Founder discovers GPT-4o can do barcode reading and package damage detection zero-shot. No training. Ship it. They are thrilled.</p><p><strong>Month 6</strong>: Product gets traction. API costs are $3,000/month. Acceptable.</p><p><strong>Month 12</strong>: Real customers onboard. Volume is 10x. API costs hit $30,000/month and climbing. Latency is inconsistent. The model provider ships an update that subtly breaks their lightning edge cases.</p><p><strong>Month 18</strong>: They hire an ML engineer to build a dedicated damage detection model. Smaller, faster, cheaper, and entirely in-house.</p><blockquote>(This is not hypothetical. Speech analytics company Convirza saw a 10x cost reduction and an 80% throughput increase by shifting from large API models to fine-tuned in-house 8B parameter models. E-commerce platform Faire achieved 28% accuracy improvements switching from GPT to custom LLMs.)</blockquote><p>Large foundational models are the gateway drug. They let founders validate ideas fast. But as companies mature, those API calls become the single largest line item on the cloud bill. The economics force them toward specialized models.</p><p>The same pattern happened with cloud infra. Startups start on AWS. Some stay forever. But as revealed, Dropbox saved $74.6M over two years by moving their core storage off AWS to custom infra. The pattern repeats. Only now it is not compute you are bringing in-house. It is intelligence.</p><h3>The Narrow Model Opportunity</h3><p>The same VLMs eating early-stage company market share are creating demand for narrow models.</p><p>VLMs are expensive, slow, and general. They hallucinate. They can not guarantee sub-pixel accuracy. They require internet connectivity. They do not run on a $500 edge device.</p><p>For a startup in month 1, none of that matters. For a startup in month 18 with real customers and real SLAs, all of it matters.</p><p>Taking a 70B parameter general model and distilling it into a 5M parameter model that runs at 120fps on a Jetson for one specific task. That is not prompt engineering. That is computer vision engineering.</p><blockquote>(Consider the economics: processing 1 million images a day through the Gemini 3 Pro API costs roughly $188,000 per year just in input tokens — turn on internal reasoning or extract hefty details, and that cost balloons further. A custom TensorRT-optimized model running on a $249 NVIDIA Jetson Orin Nano does the same volume perpetually for the cost of electricity. That margin is the business.)</blockquote><h3>CV Roles Are Evolving, Not Dying</h3><p>Traditional “train YOLO for task X” roles are declining. But they are being replaced by roles that combine vision expertise with systems thinking. Vision Systems Architects who decide when to use an API vs when to build from scratch. Edge ML Engineers who take cloud models and make them work on drones with 5W power budgets. Multimodal ML Engineers building models that fuse vision with language, depth, and action.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ll22tzSl7HgPzFb9Cn1g3Q.png" /><figcaption><em>Market demand is shifting. Real job postings from recent LinkedIn ‘Who is Hiring’ networks highlight the growing need for engineers who can build complex, constraint-based vision systems.</em></figcaption></figure><p>The role is not disappearing. It is leveling up. Fewer people building YOLO from scratch. More people building systems that know when YOLO is the right answer and when Gemini is.</p><h3>System Building is the Moat</h3><p>If your entire value proposition is “I can train a model,” you are in trouble. That skill is getting commoditized by AutoML, by VLMs, by transfer learning pipelines that any SWE can run.</p><p>But “I can build a vision system that works in production under constraints” is a different story. That means knowing when a classical approach beats a neural network. Building data flywheels. Designing fallback systems. Handling camera calibration, lens distortion, lighting variation. Making systems work at the edge, on a budget, with deterministic latency.</p><p>None of that can be replaced by an API call. It requires understanding physics, optics, compute constraints, and failure modes. It requires judgment from experience. And that experience compounds.</p><p>The general interest in AI has exploded, but the deep, constraint-based knowledge required for systems engineering hasn’t. Look at the divergence between general AI interest and Computer Vision search trends. The hype is in prompting; the moat is in building the system beneath it.</p><p>Same goes for multitask learning. Companies that can not go full E2E are building single models that handle detection, segmentation, depth, and lane detection simultaneously. The loss balancing, architecture design, evaluation frameworks. This is not something you can prompt your way through. It requires deep vision knowledge about how detection and depth estimation share features and where they compete. So it is worth sharpening your skills in this area if you are considering a career in computer vision or planning to enter the workforce soon.</p><h3>The Bottom Line</h3><p>The career is not a dead end. It is a fork. The old path (train single purpose YOLO, deploy MobileNet, repeat) is narrowing. The new path (build vision systems, own the edge, know when NOT to use a foundation model) has more demand than ever.</p><p>The same forces killing traditional CV roles are simultaneously creating higher-value ones. More startups, more niches, more need for specialized, efficient, production-ready vision. The job title might change. The LinkedIn keyword might shift from “object detection” to “vision systems” or “multimodal ML.” But the work, making machines see accurately under real-world constraints, is not going anywhere.</p><p>The water is rising. But so is the ground we are standing on.</p><p><em>Disclaimer: AI was used in editing and generating graphics included above.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6c88f572011" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>