<?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 Altuna Akalin on Medium]]></title>
        <description><![CDATA[Stories by Altuna Akalin on Medium]]></description>
        <link>https://medium.com/@aakalin?source=rss-c4165496d984------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*KusbdUJhNdDWrM4IE6q9SA.jpeg</url>
            <title>Stories by Altuna Akalin on Medium</title>
            <link>https://medium.com/@aakalin?source=rss-c4165496d984------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 11:26:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aakalin/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[Why we must develop and practice meta-skills in the age of AI]]></title>
            <link>https://medium.com/data-science-collective/why-we-must-develop-and-practice-meta-skills-in-the-age-of-ai-7d719a931599?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d719a931599</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 12:15:29 GMT</pubDate>
            <atom:updated>2026-11-04T12:15:29.574Z</atom:updated>
            <content:encoded><![CDATA[<p>If you write code or analyze data, you are living through the greatest cognitive offloading experiment in human history. But away from the headlines, a much more personal crisis is playing out on our monitors. It’s happening in code reviews, design docs, and troubleshooting sessions: <strong>our core mental muscles are starting to rust. Or worse, for the younger generation, these muscles may never develop as AI removes friction and obfuscates process details.</strong></p><p>As someone who builds AI tools and works in biomedical data science, I see both sides of the coin. I love the support these tools provide. But I am also aware that our minds can fall into a trap, what aviation engineers call <strong>automation complacency</strong>. When a pilot relies on autopilot for too long, manual flying skills atrophy. More dangerously, pilots lose their internal “mental model” of the system, making them helpless the moment something unexpected goes wrong.</p><p>When you remove all the friction and frustration from technical work, you remove the very resistance that builds your intuition and meta-skills.</p><h3>Domain experts benefit more from the AI</h3><p>Automation doesn’t affect everyone equally and this puzzled me for a while. I assumed that once AI agents capable of assisting with code arrived, everyone in our lab would become infinitely more productive. But productivity didn’t rise across the board, even though everyone adopted agentic AI in some form. The seniors in the lab consistently obtained far more value from the AI tools than juniors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7qloFKgP2cux-ApR.png" /><figcaption>“AI equalizes performance on well-structured, routine tasks where the “correct” approach is well-represented … But on complex tasks requiring deep domain judgment, contextual reasoning, and iterative refinement AI functions as a cognitive amplifier that widens rather than narrows the expertise gap.” An T. <em>HHAI 2026 </em>doi:<a href="https://doi.org/10.3233/FAIA260506">10.3233/FAIA260506</a></figcaption></figure><p>The clues lie in recent findings by <a href="https://doi.org/10.1287/orsc.2025.21838">Dell’Acqua et al.</a> and <a href="https://ebooks.iospress.nl/doi/10.3233/FAIA260506">Tao An</a>. What those papers argue is that the AI has a context specific effect on output quality. It depends on task complexity and user’s domain expertise. For low complexity tasks, it acts as an equalizer. Now everyone can make a nice figure or a web page just by prompting. But for high-complexity tasks, experts benefit more. Why is that though?</p><p>Research work wether wet-lab or computational needs experimentation and cycling through ideas that are born out of failures or past experiences. Pre-AI, we had time, expertise and intuition to explore only a few ideas. With AI, we can explore far more ideas or paths regarding a research problem. However, <strong>being able to explore more doesn’t mean you know where to go</strong>. You need the domain expertise and intuition, possibly build through failures, to steer your research to productive paths.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2Kogi1OK9agutPde.png" /><figcaption>AI expands exploration, not necessarily productivity: Without AI, few paths are practical to explore. With AI, exploration becomes easier but one can also get stuck more frequently in non-productive paths if domain knowledge to steer the AI is lacking.</figcaption></figure><p>Seniors already possess the internalized mental models needed to decompose problems, steer the models, and spot subtle hallucinations or drift from initial goals; juniors possibly lack that reference frame. AI has taken over the routine mechanics of coding. What’s left are the <strong>meta-skills</strong>: the higher-level abilities that govern how we think, how we come up with tests, formulate hypotheses, and know when a system is lying to us.</p><h3>The recruitment test born out of necessity</h3><p>If meta-skills are the only differentiator that matters, traditional coding challenges that we used to send to candidates are obsolete. AI solves them instantly. That’s why I had to invent a test for raw reasoning and critical thinking.</p><p>The mechanics of the test were trivial: You provide input variables and you observe one output. To solve it, a candidate had to reverse-engineer the hidden mathematical rule by running tests and observing the responses. I assumed experienced engineers, data scientists, and sharp graduates would breeze through it. However, many smart candidates repeatedly failed in predictable ways: changing multiple variables at once, ignoring edge cases, and guessing from memory rather than tracking results.</p><p>When we outsource our thinking to AI that fails with fluent confidence, we will forget how to systematically interrogate a black box. If systematic reasoning is an atrophied skill, how do we train it when software removes all friction?</p><h3>Three rules for “deliberate friction”</h3><p>If we want to stay sovereign over our work instead of becoming passive passengers, throwing away our tools isn’t the answer. We have to introduce deliberate friction into our training and daily habits. Just as physical workouts became essential once manual labor was automated, we now need cognitive calisthenics to keep our technical meta-skills sharp.</p><p>Here is how you can practice it starting today:</p><h4>1. Start with a blank page</h4><p>Never hand a raw, unstructured problem straight to an AI prompt. Before you open a screen or type a query, spend 10 minutes with a physical pen and a piece of paper. Sketch out your hypothesis, map the state transitions, or write down what you <em>think</em> is causing the bug. The physical act of structuring a thought before a machine touches it builds the neural pathway required to audit the machine’s output later. You cannot edit what you do not understand.</p><h4>2. Probe the boundaries, don’t take the easiest path</h4><p>When troubleshooting an unfamiliar system, be it an AI-generated script, or a third-party report, humans usually only test what confirms our biases. We check if it works for standard inputs. An adversarial thinker does the opposite. You must actively hunt for the breaking point: <em>What happens at zero? What happens with negative values? What if I inject a decoy variable?</em> True understanding of a system lives at the edges, not in the middle.</p><h4>3. Own the hypothesis, outsource the legwork</h4><p>Let AI handle the mechanical tasks such as formatting data, scaffolding boilerplate, scraping documentation, or translating your rough logic into code. But <strong>never outsource hypothesis formation or verification to AI.</strong> If the machine writes both the core implementation and the tests, you are living in a closed loop of hallucinated confidence.</p><h3>Test your own skills</h3><p>Reading about cognitive fitness is like reading about lifting weights, it doesn’t build any muscle unless you do the reps.</p><p>To see how your own troubleshooting meta-skills hold up under pressure, I’ve adapted our black-box recruitment diagnostic into a short, 3-round interactive challenge called <strong>Beat the Black Box</strong>.</p><p>Grab a pen and paper, see if you can make it through.</p><p>👉 <a href="https://blackboxgame.web.app/"><strong>[Play the “Beat the Black Box” here]</strong></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L8OFeqrZohb6uFFI.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d719a931599" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science-collective/why-we-must-develop-and-practice-meta-skills-in-the-age-of-ai-7d719a931599">Why we must develop and practice meta-skills in the age of AI</a> was originally published in <a href="https://medium.com/data-science-collective">Data Science Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Towards Deterministic and Reproducible Agentic Data Analysis]]></title>
            <link>https://pub.towardsai.net/towards-deterministic-and-reproducible-agentic-data-analysis-778d8ec3a89d?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/778d8ec3a89d</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Wed, 09 Sep 2026 19:01:01 GMT</pubDate>
            <atom:updated>2026-09-09T19:01:01.462Z</atom:updated>
            <content:encoded><![CDATA[<p>People of all ages and experience levels who know a little bit about LLMs and want to do data analysis without coding are turning to agentic tools. These tools are numerous by now and include tools from model builders such as Anthropic (Claude Code) and OpenAI (GPT Codex), as well as many other open-source alternatives (OpenCode, etc.).</p><p>What is an agent in this context? An agent is usually a language model most of the time with vision capability (can see images and convert them to text). It can inspect data, install libraries, write scripts, troubleshoot failures, and produce an analysis. This can all be done using only the user’s instructions. However, every request to the agent will result in a new batch of text generation and potential web searches, and a new batch of code will be created and executed. This new code will not be identical to the previous runs even if you provided the same task. This stochasticity is inherent in large language models, and we have to live with it. But this hurts reproducibility, reusability, and racks up the costs. For example, if you want to perform a routine analysis of transcriptome data, there is not much variation in the code you will use. You or your agent don’t have to reinvent the wheel every time you want to analyze data. [<strong><em>SPOILER</em></strong>: if you don’t want to read further here is how to make agents more reproducible: <a href="https://github.com/rolv-io/recipe">https://github.com/rolv-io/recipe</a>]</p><h3>How to deal with stochasticity in agentic coding?</h3><p>Luckily, there are mechanisms to provide pre-determined instructions to the agents so that they don’t have to reinvent things every time we give a task that requires specialized knowledge or skills.</p><p>SKILL.md files and their accompanying scripts, references, and other files are the prevailing way to provide these instructions. They allow us to package instructions and working code so that an agent does not need to reinvent an entire workflow every time.</p><h3>Current problems with agentic coding and skills</h3><p>However, reusable instructions provided with the skills approach are not the same thing as a reproducible and continuously improving workflow. Here are the current problems with the agentic coding and skills approach.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JVv7fJZCD_CqsRSi.png" /></figure><h4>1. Skills reduce repeated reasoning, but agents still make unnecessary decisions</h4><p>A good skill can tell an agent how a task should be performed and provide scripts that implement established parts of the process. This already saves tokens and reduces reinvention.</p><p>However, whenever the skill leaves a decision open, the agent has to reason about it again. It may choose a different order of operations, a different library, a different intermediate representation, or a different response to an error.</p><h4>2. Skills describe the intended workflow, but do not fully define its execution</h4><p>A SKILL.md file can say:</p><blockquote><em>Load the data, validate it, transform it, run the analysis, and generate the output.</em></blockquote><p>It can also provide scripts for doing each of these things.</p><p>But there is still a difference between describing a workflow and explicitly representing its execution graph.</p><h4>3. Troubleshooting can destroy parts of the workflow that were already working</h4><p>This becomes particularly important when something fails. An agent trying to solve a problem may modify several scripts, change a dependency, alter an earlier transformation, or rewrite an important part of the workflow simply because doing so seems plausible. I have experienced this many times, even with frontier models. The agents were trying to fix problems at the wrong steps.</p><h4>4. Dependencies should be mandatory, not optional</h4><p>Skills can specify dependencies and include environment files, but they do not necessarily have to. A skill may work because the required libraries, tools, or system dependencies happen to already be available in the agent’s environment. Dependency management is critical for reproducibility and should not be optional.</p><h4>5. Skills preserve solutions, but troubleshooting knowledge is still easy to lose</h4><p>The initial skill may describe the correct workflow. However, the more interesting knowledge often appears later. Perhaps a particular CSV export occasionally changes its encoding. An API starts returning a slightly different schema. A package version introduces a bug. Packages have different installation procedures in windows vs linux. A model produces malformed output under a particular condition. An apparently harmless transformation turns out to break one class of datasets.</p><p>Agents solve these problems during execution, but those discoveries should not disappear with the session. The problems they encounter are usually not one-off problems.</p><h4>6. A successful skill is still not necessarily a correct skill</h4><p>Repeated successful execution can create false confidence. A workflow might run without errors while quietly producing incorrect results. What is worse, it may be creating imaginary datasets to make certain parts work. An expert who is lazy not to check the details or a novice who has no idea how thigns actually work will miss these problems easily. No mechanism is forcing skill builders to provide tests with their skills.</p><h4>7. Sharing skills is not the same as learning from executions</h4><p>A shared skill lets many agents start from the same knowledge.</p><p>But there is another step: allowing the shared workflow to improve from what those agents discover while using it. This is a completely missed opportunity. Agents are again solving problems over and over while using faulty skills, and their learnings are never transfered to other agents. There is no hive-mind, although maybe there should be.</p><h4>8. Shared skills introduce a new security problem</h4><p>Once a skill contains scripts, tools, dependencies, instructions, and contributions from many agents or people, it becomes executable infrastructure.</p><p>That creates a supply-chain problem. The skills can include suspicious dependencies, malicious commands, unexpected network access, dangerous tools, credential access, prompt-injection-like instructions, or changes that unexpectedly expand what the skill is allowed to do previously. There is no central check on what security risk each skill can pose.</p><h4>9. Important constraints can lose their context over time</h4><p>As workflows evolve, some steps, dependency pins, or troubleshooting rules may begin to look unnecessary. But they may exist because of failures discovered in previous executions.</p><p>If the reason for a constraint is not preserved, future agents may remove or change it without understanding why it was introduced, causing previously solved problems to reappear.</p><h3>Proposal to remedy these issues: recipe.md</h3><p>The idea is to extend the current skill approach with a more structured artifact: a recipe.md file and accompanying files that define not only how a task should be approached, but how an established workflow should be executed, maintained, tested, and improved over time. In hindsight, we can call this skill files or skill.md as well, but the divergence from a regular skill file set is substantial and we don’t want people to think this is same as the skill files.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*s-b7P7zZ2NqwKcNU.png" /></figure><p>A recipe would add several things that are not mandatory in today’s skill approach.</p><h4>1. A deterministic execution graph</h4><p>The recipe should explicitly define the steps of the workflow, their order, their inputs and outputs, and the conditions under which different branches can be taken.</p><p>Instead of repeatedly asking the agent to decide how to perform an established task, the recipe defines the default execution path:</p><blockquote><em>Step A → validate → Step B → Step C → validate output</em></blockquote><p>The agent can still reason where judgment is required, but it should not redesign parts of the workflow that have already been established.</p><p>This should reduce unnecessary token use and make repeated executions more consistent. This idea is of course very similar to workflow managers such as snakemake, nextflow, apache airflow etc. The fundamental difference is that AI runs, troubleshoots and intreprent the steps. The artifacts are human readable rather than a programming language.</p><h4>2. Troubleshooting rules for every step</h4><p>When something fails, agents should not have unlimited freedom to modify the entire workflow.</p><p>Each step or component can carry a troubleshooting policy such as:</p><ul><li>normal : reasonable to modify while troubleshooting;</li><li>core (or touch-last): known to be stable and should only be changed after more likely causes have been investigated;</li><li>do-not-touch: should not be changed automatically. This should be used sparringly but probably has some use cases.</li></ul><p>This allows knowledge from previous executions to constrain future troubleshooting.</p><p>A recipe also specifies via the workflow graph which upstream and downstream steps depend on the component being modified, so the agent can understand the likely consequences of a change before making it.</p><p>Preserve why important rules exist</p><p>Recipes should record not only important constraints but also a short explanation of why they exist. If a step is marked core, the recipe should say why. If a validation step was introduced because of a previous failure, that relationship should remain visible.</p><p>The detailed evidence can remain in the troubleshooting record (which will be explained below), while the current recipe contains a short rationale and a reference to the relevant incident.</p><p>For example:</p><pre>step: normalize_columns<br>troubleshooting_policy: core<br>reason: Previous changes to this step caused downstream schema mismatches.</pre><p>This prevents future agents from removing safeguards simply because they no longer know why those safeguards were introduced.</p><h4>3. Mandatory dependency declaration</h4><p>Unlike skills, where dependency information may or may not be included, a recipe should not be considered complete unless its dependencies are explicitly declared.</p><p>This can include libraries and versions, command-line tools, external services, APIs, models, system dependencies, and dependencies between individual workflow steps.</p><p>The goal is not only to document what is needed, but to make it possible to reproduce a known-working environment and quickly identify dependency changes when something stops working.</p><h4>4. A persistent troubleshooting record</h4><p>Every execution that encounters a meaningful problem should be able to add structured information to the recipe’s troubleshooting history.</p><p>That record might include:</p><ul><li>what failed;</li><li>the symptoms or error;</li><li>the likely diagnosis;</li><li>what the agent tried;</li><li>which attempts did not work;</li><li>what eventually fixed the problem;</li><li>which files or dependencies changed;</li><li>whether the solution passed the recipe’s tests.</li></ul><p>Future agents should consult this record before beginning their own troubleshooting.</p><p>The aim is simple: if one agent has already spent time diagnosing and solving a recurring problem, another agent should not have to rediscover the same solution from scratch.</p><h4>5. Automated tests on known data</h4><p>A recipe that executes successfully is not necessarily producing the correct result.</p><p>Recipes should therefore be able to include automated tests, including small toy datasets for which the expected outputs are already known.</p><p>After an agent changes a script, dependency, or workflow step, the recipe can be run against these examples to verify that previously correct behavior has not been broken.</p><p>This also provides a safer way to incorporate improvements discovered during troubleshooting.</p><h4>6. Learning across recipe executions and changes</h4><p>A shared recipe should improve from the experience of the agents using it.</p><p>After an execution, an agent can submit what it learned: a newly discovered failure mode, a better troubleshooting procedure, an important dependency constraint, or an improvement to an existing step.</p><p>These contributions should not necessarily be incorporated automatically. They can first be tested, compared with the existing recipe, scanned for security issues, and, when necessary, approved by a human.</p><p>Once accepted, however, the improvement becomes available to every subsequent execution of the recipe.</p><p>The important difference is that knowledge flows back into the shared workflow rather than remaining inside an individual agent session.</p><h4>7. Central security scanning</h4><p>Recipes and their accompanying files are executable infrastructure and should be treated accordingly.</p><p>A central repository can scan recipes for malicious or suspicious behavior before they are distributed or updated. This could include unexpected network access, suspicious package dependencies, credential access, dangerous shell commands, malicious tools, prompt-injection-like instructions, or changes that substantially expand the permissions of an existing recipe.</p><p>The same checks can be applied whenever an agent proposes an update to a shared recipe.</p><h4>8. Versioned and human-governed improvement</h4><p>Recipes should be versioned so that changes to workflow logic, dependencies, troubleshooting rules, and tests are visible and reversible.</p><p>The initial recipe should be reviewed by a human, and important changes can go through the same process. The purpose of human oversight is to approve the reusable procedure and significant changes to it, rather than supervise every individual run.</p><p>A version change can then capture both <strong>what changed</strong> and <strong>why</strong>, while the troubleshooting record preserves the detailed history behind that change.</p><h3>The resulting lifecycle</h3><p>The resulting process would look something like this:</p><ol><li><strong>execute recipe</strong></li><li><strong>encounter a problem</strong></li><li><strong>consult previous troubleshooting</strong></li><li><strong>troubleshoot within defined boundaries</strong></li><li><strong>test the solution</strong></li><li><strong>record what was learned</strong></li><li><strong>submit the improvement</strong></li><li><strong>scan and review</strong></li><li><strong>update the shared recipe</strong></li></ol><p>In this model, recipe.md adds a stricter layer for workflows that are already established and where determinism, reproducibility, testing, and accumulated operational knowledge matter.</p><p>The goal is not to eliminate agent reasoning. It is to avoid spending agent reasoning on decisions that have already been made and problems that have already been solved. Most of these ideas are implemented in our app <a href="https://rolv.io/">rolv.io</a> (a local-first agentic data analysis app), and the formal definition is here: <a href="https://github.com/rolv-io/recipe">https://github.com/rolv-io/recipe</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=778d8ec3a89d" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/towards-deterministic-and-reproducible-agentic-data-analysis-778d8ec3a89d">Towards Deterministic and Reproducible Agentic Data Analysis</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why data science needs more than an agentic terminal]]></title>
            <link>https://medium.com/design-bootcamp/why-data-science-needs-more-than-an-agentic-terminal-90738c642fde?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/90738c642fde</guid>
            <category><![CDATA[science]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 00:18:31 GMT</pubDate>
            <atom:updated>2026-09-05T00:18:31.917Z</atom:updated>
            <content:encoded><![CDATA[<h4>The problem with general AI coding assistants in science</h4><p>Over the past year, tools like Claude Code, Cursor, and Copilot have transformed software engineering and also data analysis. You can point them at an API backend, and they write boilerplate with incredible speed.</p><p>Yet, when life scientists and computational biologists try using generic coding agents for single-cell RNA-seq, clinical trial analysis, GWAS, or metabolomic<strong>s</strong>, things quickly become complicated for novices and experienced practitioners alike:</p><p>1. <strong>The “spaghetti script” trap</strong>: Terminal agents generate endless chronological code blocks. Soon, nobody knows which variable came from where, which intermediate file is current, or how to reproduce the figure.</p><p>2. <strong>Scientific data and concept blindness</strong>: Generic LLMs know GitHub and they can search the web, but they are blind to or not directed to NCBI GEO, PubMed literature, ClinicalTrials.gov, or patent databases.</p><p>3. <strong>The compute requirements are different</strong>: Scientific data doesn’t live in a 20-line CSV on your desktop or laptop. It lives on an HPC cluster as 50 GB BAM/RDS files. A local terminal assistant cannot seamlessly orchestrate detached SSH jobs while keeping track of your outputs. Handling this situation needs extra care and provenance</p><p>4. <strong>Re-inventing the wheel</strong>: Every Claude Code or Cursor session starts with a blank slate, hallucinating ad-hoc scripts rather than using verified and trusted protocols. Skill files help, but they are not perfect and far from providing the reproducibility that scientific data analysis requires.</p><blockquote><em>Scientific data analysis doesn’t just need code generation. It needs structured workflow execution, provenance, reproducibility, and deep domain intelligence.</em></blockquote><p>That is why we keep building <strong>rolv</strong>, and happy to introduce v3.</p><h3>Rolv: The agentic platform for life sciences</h3><p>Rolv is a local-first desktop application that bridges the gap between exploratory AI chat and visual data analysis workflows. Below, we are introducing some of the key features of this system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lAK5Q0x6pJ4TyGK9.png" /><figcaption>Rolv chat and visual DAG interfaces</figcaption></figure><p>1. From AI chat notebook to visual graph for workflows</p><p>The data analysis usually starts messy: inspecting columns, trying different normalization methods, filtering outliers. You need a few quick tools to get started, and you want to deal with as few complications as possible in terms of installing packages, reading data, and gathering basics about the content and quality.</p><p>With Rolv, you can explore datasets freely in chat. Once your logic works, click ”<strong>Step into Graph</strong>”. Rolv’s Architect agent reads your conversation history and synthesizes a clean, modular<strong> Directed Acyclic Graph (DAG) </strong>where each step (Quality Control, Normalization, Clustering, Differential Expression) becomes an isolated canvas node with defined inputs and outputs. Every node runs in an isolated sandbox with clear input &amp; output contracts. We have smart invalidation and staleness propagation in the workflow graph: Tweak step 2, and Rolv only re-runs downstream dependencies, leaving your upstream compute. You can also pin long-running jobs not to re-run when your workflow changes.</p><p>If you want, you can entirely skip the chat and start building your workflow node by node with AI assistance. You can let the AI code or you do it yourself and ask AI to fix mistakes.</p><h4>2. Built-in biomedical research &amp; multi-modal vision</h4><p>We don’t want <em>rolv</em> just to write code. Ideally, it should reason over the results with scientific context and should have the tools to gather that context. Here are some examples of the tools and sub-agents where the agent use to make sense of the output of data analysis.</p><p><strong>NCBI GEO &amp; PubMed Native Search</strong>: Ask Rolv to ”find public GEO datasets for KRAS-mutant pancreatic cancer and compare with our cohort”. The agent searches GEO accessions and PubMed literature directly, extracting PMIDs, abstracts, and metadata into downstream compute steps.</p><p><strong>Plot vision</strong>: When generating a volcano or UMAP plot, Rolv converts the output to high-resolution images and visually inspects the graph to catch label overlaps, skewed axes, or outlier artifacts before you even notice them.</p><p><strong>Structured research nodes</strong>: Transform unstructured literature and clinical trial searches into structured JSON tables and synthesis reports that feed directly into analysis code.</p><h4>3. Near-bulletproof reproducibility</h4><p>In standard computational notebooks and AI chat tools, reproducibility can rarely be assumed. Most scientific papers you will encounter will have some code or even python notebooks however, the software environment that produced the plots may not be documented or even reproduced. In addition, you will have a bunch of scripts it won’t be immediately clear how things are run and what output produced which plot. There is usually no provenance.</p><p>Rolv enforces computational reproducibility by design:</p><p>1. Process-level node isolation: Every node runs in a fresh, isolated subprocess. Nodes pass data through explicit disk contracts (`input.json` / `output.json`), never shared in-memory variables.</p><p>2. Automated provenance ledger: Every plot, table, and RDS object is cataloged in `output.json` with the exact producing node ID, parameters, and timestamps.</p><p>3. Environment locks: Workflows auto-generate env/python-requirements.txt and env/r-packages.txt. Rolv provisions isolated Mamba environments on demand so your pipeline runs (almost) identically on a colleague’s machine.</p><p>4. One-click supplementary bundles: Export your entire workflow as a self-contained, versioned Recipe package** ready for peer-review submission, GitHub release, or institutional sharing.</p><h4>4. Hybrid Compute: Laptop to remote computer</h4><p>Biological datasets quickly exceed local RAM. Rolv features a hybrid execution possibility to deal with out-of-memory data analysis situations:</p><ul><li>Right-click any node in your visual workflow to toggle execution between <strong>Local Laptop/Desktop (Apple Silicon / PC)</strong> and <strong>Remote Lab Server / HPC (via SSH)</strong>.</li><li><strong>Detached background execution</strong>: Heavy compute jobs run autonomously on your remote server. Close your laptop, go home, and open Rolv the next day — the app seamlessly reconnects to the detached job, streams remaining logs, and downloads the generated figures.</li><li><strong>Safeguard against credential leaks</strong>: Passphrases and SSH keys are encrypted directly in your own system.</li></ul><h4>5. Recipes &amp; institutional memory: Never start from scratch</h4><p>Why ask an AI to invent a DESeq2 pipeline from scratch when your lab already has a golden standard?</p><p>Rolv introduces the concept of <strong>recipes</strong>. These are modular, shareable packages that bundle workflow DAGs, documentation, environment locks, and scripts. They are not same as “skills”. The recipes require more structure, clear instructions, and provenance. Imagine a cook who has the skills to craft a perfect meal but doesn’t write his creations as a recipe. Every time she cooks the same meal, they differ sometimes slightly and sometimes greatly. This is not what we want in science, and in a good meal we want to enjoy repeatedly.</p><ul><li><strong>Local recipe database</strong>: Rolv indexes your lab’s recipes in an on-device database. When you describe a task, the agent automatically retrieves and adapts verified blueprints.</li><li><strong>Import existing pipelines</strong>: Have legacy workflows? Rolv automatically imports Snakemake and Nextflow pipelines, translating rule blocks into interactive visual canvas nodes.</li><li><strong>Collaborative changes to recipes:</strong> You can share recipes to the community, propose updates with automated AI risk assessments, and maintain reproducible institutional standards. This creates a centralized intelligence where one agent’s learnings and skills are propagated to the central repository.</li></ul><h4>6. “Touch-last” auto-healing: Fixing errors without breaking core functionality</h4><p>When an analysis script throws an error, conventional AI agents rewrite everything in sight, often breaking upstream calculations.</p><p>Rolv introduces “<strong>touch-last troubleshooting” </strong>policies:</p><ul><li>Mark critical statistical models or normalization algorithms as <strong>touch-last</strong>.</li><li>When errors occur downstream, Rolv’s auto-healing engine prioritizes patching wrapper code and input formatting rather than touching your core validated math.</li><li><strong>Healability Gating</strong>: The agent knows the difference between a real code bug and a system issue (missing package, disk full, network timeout), preventing infinite self-heal loops.</li></ul><h4>7. Local-first, privacy-first, zero-cost BYOK</h4><p>The data and discoveries are your most sensitive assets. This is why rolv is built around local and privacy-first scenarios.</p><ul><li>Rolv stores your projects, data, and recipes 100% locally.</li><li>Although the rolv app provides fast online LLM models by default, you can easily bring your own LLM. Connect local Ollama models (e.g. llama3, qwen, embeddinggemma) for 100% free, offline, air-gapped execution, or plug in Groq, OpenRouter, or LLM models provided by your own institute/company.</li></ul><h4>See the Difference: rolv vs. traditional coding assistants</h4><p>Here is a feature comparison table for rolv vs other agentic frameworks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DemIYy1f6S_DmmGX.png" /></figure><h3>How to get started</h3><p>Science moves too fast to wrestle with fragmented scripts and broken Jupyter notebook states and environments. Try out a workflow platform built specifically for the rigor, scale, and complexity of modern scientific discovery.</p><ul><li>🚀 Download Rolv for macOS, Windows, and Linux: <a href="https://github.com/rolv-io/rolv-releases">https://github.com/rolv-io/rolv-releases</a></li><li>📖 Explore the Documentation: All the documentation is within the app. Either ask a direct question about features in the chat or read the in-app documentation.</li><li>💬 Join the Community &amp; Share Recipes: Make a recipe and submit it for community review</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90738c642fde" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/why-data-science-needs-more-than-an-agentic-terminal-90738c642fde">Why data science needs more than an agentic terminal</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to build your own AI peer-review simulation]]></title>
            <link>https://briefs.aiadvances.org/how-to-build-your-own-ai-peer-review-simulation-222b93770f72?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/222b93770f72</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Thu, 03 Sep 2026 11:52:34 GMT</pubDate>
            <atom:updated>2026-09-03T11:52:34.150Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*P2d1QYZvoTSXtHsO" /><figcaption>Photo by <a href="https://unsplash.com/@guerrillabuzz?utm_source=medium&amp;utm_medium=referral">GuerrillaBuzz</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Before submitting a paper to a journal, it can be useful to put it through a simulated round of peer review first. An AI peer-review simulator can help you uncover unclear arguments, methodological weaknesses, overstated claims, and the kinds of questions a real reviewer might ask. Below, we will build such a workflow and submit one of our pre-prints to it and see what happens.</p><p>Before we start, <strong>here is the final </strong><a href="https://github.com/al2na/peer_review_automation"><strong>workflow</strong></a><strong>.</strong> You can download and use it already if you don’t have the time to read the rest of the post.</p><p>So, the questions I have in mind are: Can an AI system produce useful reviewer comments? Will it identify genuine weaknesses in the paper? And, perhaps more importantly, can we use those comments to improve a manuscript before sending it to a journal?</p><p>Let’s build it and find out.</p><h4>The general idea</h4><p>The general idea is that we will build a workflow starting with claims extraction. Those claims will be matched to supporting results and methods. A literature review will assess novelty and precedent for the results and methods. All of these steps will end up in a final synthesis step that will produce major and minor points to address. Each task will be represented as node in the user interface and we would be able to control its behaviour via instructions. <strong>Important caveat</strong>: Many steps are generative, meaning that re-running the same step with the same input might result in different output. It is a good idea to run critical steps multiple times.</p><h4>Privacy enabling (IMPORTANT)</h4><p>Before we start explaining the parts of the workflow, we should talk about how to enable privacy. For each node, you can enforce a local LLM model to be used. In this case, no cloud-based LLM models will be used to perform the tasks, but depending on your setup, you will still perform a PubMed or web search using terms derived from your input text. You can also completely turn off any web-based tool for search and only rely on LLMs reasoning skills to carry out the sub-tasks. Turning off PubMed and/or web searches will likely substantially change the results and increase hallucinations, but the option is always available if you want to use only the local LLM on your device. In addition, running the local models will usually be slower if you don’t have a good GPU-model combination.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/508/0*5R8sZL066tGOQdma.png" /></figure><h4>Input node</h4><p>Input node defines that we will feed in PDF, and the 2ndOpinion app will extract the text and text representation of images in the PDF. You can also modify this node to accept text and paste your text here.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GgQXIknDBfLj4Fgz.png" /></figure><h4>Claim extractor</h4><p>This node/task will extract the main claims from the text. It will extract top 3 main claims that are defined by a few criteria in the instructions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3YKJNEUv81NSzVp_.png" /></figure><h4>Establish support needed for claims</h4><p>For each claim, establish what kind of experiments/results/analysis needed to support the claim. Again, there is a large set of instructions, but we are looking potential evidence for the claims to be supported.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1ER-FHX3hN6ds0Ui.png" /></figure><h4>Extract support provided in the manuscript for a given claim</h4><p>For a given claim and the text, we are looking for supporting evidence for the claim in the text. This bit outputs the supporting evidence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ymfC2qyX5kQ8hGIN.png" /></figure><h4>Search literature for relevant articles</h4><p>This bit helps us see if the claims are novel and/or plausible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GxBt1RK_RPrNqNh-.png" /></figure><h4>Compare expected support with provided support for the claims</h4><p>This bit checks if the provided support for a claim matches what is expected. Remember, we extracted the expected support earlier, which generally provides an extensive list of what would be needed for the claim to be supported.</p><h4>Final synthesis</h4><p>Every important bit is fed into a final synthesis task and this gives major/minor concerns to address.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8k3QPX2-fQ0_EVxI.png" /></figure><h4>How did the workflow do?</h4><p>We ran the workflow on <a href="https://www.biorxiv.org/content/10.64898/2026.06.18.733198v1">one of our latest preprints.</a> The results confirmed what we already know: we need more validation and methodological details. Full output is <a href="https://github.com/al2na/peer_review_automation/blob/main/Paper_peer_review_Execution_Report_Report.md">here</a>. One issue is partly that the the web search finds our paper already, so making that part less useful than it can be. I think this can be adjusted with some prompt engineering but I’m out of time for this experiment. Someone else try and let me know.</p><h4>How to improve this further</h4><ul><li>Add stability to claim extraction and pubmed searches by doing them multiple times. You can add nodes with same tasks and an additional syhtnesis tasks to use the consesus output.</li><li>You can also add nodes/tasks that comment on writing clarity and point out issues there.</li><li>You can try to modify literature search instructing to ignore identical research.</li><li>Anything else you can think of comment below or whereever you see this post.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=222b93770f72" width="1" height="1" alt=""><hr><p><a href="https://briefs.aiadvances.org/how-to-build-your-own-ai-peer-review-simulation-222b93770f72">How to build your own AI peer-review simulation</a> was originally published in <a href="https://briefs.aiadvances.org/">AI Advances Briefs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The private, multimodal AI agent platform for clinical & scientific research]]></title>
            <link>https://aakalin.medium.com/the-private-multimodal-ai-agent-platform-for-clinical-scientific-research-339ab60230b2?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/339ab60230b2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[medicine]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Thu, 27 Aug 2026 15:36:31 GMT</pubDate>
            <atom:updated>2026-08-27T19:37:25.922Z</atom:updated>
            <content:encoded><![CDATA[<p>Healthcare and biomedical research increasingly depend on AI to make sense of large volumes of scientific literature, clinical documents, reports, images, and other multimodal data. Yet many of the most capable AI systems today rely heavily on cloud infrastructure. This creates legitimate concerns around <strong>data privacy, confidentiality, regulatory compliance, and control over sensitive information</strong>.</p><p>Therefore, there is a need for a privacy-conscious desktop application designed specifically for clinicians, medical researchers, scientists, and healthcare organizations. Our new app 2ndOpinion bridges the gap between deep biomedical evidence synthesis, multimodal document intelligence, and privacy-preserving local AI execution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*koNORCjK9TR_yXIw.png" /></figure><h3>Why did we build 2ndOpinion desktop?</h3><p>In clinical practice and medical research, decision-making demands precision, source attribution, and absolute privacy. It is also essential that clinicians and researchers have a platform to test and build multi-modal AI workflows for day-to-day tasks.</p><p>Standard cloud AI models and frameworks often present data privacy concerns when handling confidential patient information. Some lack native support for multimodal patient documents (like X-rays, ECGs, and lab reports). In addition, determining which medical and scientific source material is needed for specific tasks is something that absolutely should be handled by domain experts rather than relying on web search to find relevant documents.</p><p><strong>2ndOpinion Desktop</strong> solves these issues by putting an autonomous, multi-agent AI engine right on your workstation — with the capability to run 100% locally when offline, or connect to specialized biomedical search engines and cloud workflows when authorized. It enables users to build/share custom local knowledge bases and agentic workflows.</p><h3>Key features &amp; highlights</h3><h4>1. Build your own knowledge base and let agents use it</h4><p>Today, many medical specialties rely on clinical guidelines and research papers to guide patient care. Because these guidelines and Standard Operating Procedures (SOPs) vary by country and institution, building a reliable knowledge base for medical agents is becoming increasingly important.</p><p>2ndOpinion lets users build such knowledge bases to be used by agents. These databases are completely local and can be shared with other users.</p><h4>2. Privacy-first Local AI &amp; flexible LLM usage</h4><p>2ndOpinion can be used <strong>100% offline,</strong> or it can selectively run certain tasks online and others offline. It has an intelligent fallback routing system, it can:</p><ul><li>F<strong>orce local mode:</strong> Complete air-gapped privacy via on-device models.</li><li><strong>Use </strong>c<strong>ustom API keys</strong>: Route queries directly, ideally to your institute’s own LLM deployment. You can also use your own OpenAI or Anthropic API accounts, but that doesn’t help with privacy.</li><li><strong>Use 2ndOpinion Cloud Engine:</strong> Offload heavy RAG &amp; multi-step biomedical synthesis to the 2ndOpinion Cloud backend with automatic fallback to your local model if offline.</li></ul><p>The app also comes with a <strong>local de-identification protocol.</strong> Clinical queries frequently contain sensitive patient information. In all the interfaces provided by the app, you can turn on the <strong>de-identification protocol </strong>to scrub personally identifiable information from your queries locally.<strong> </strong>This way, even if you use online models and APIs, personal information will not leave your computer.</p><h4>3. Autonomous biomedical search agents</h4><p>2ndOpinion comes equipped with pre-built, specialized tools designed for medical literature and clinical trial discovery. It can autonomously decide which resource or resources would be the best to accomplish a task and then use them. The agens have the capability to search pubmed, clinical trials, the web and read full text documents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6ikhDnppwt1M0tIf.png" /><figcaption>Part of available agents in 2ndOpinion desktop</figcaption></figure><h4>4. Custom automations (multi-step workflows)</h4><p>You can create, customize, and execute complex multi-step workflows using an intuitive canvas.</p><ul><li><strong>Graph/node Canvas:</strong> Connect input, decision/branching and agent execution nodes visually. Assign policies and capabilities to each node and let them carry out their tasks.</li><li><strong>Granular Model Policies</strong>: Control privacy on a per-node level (`local_only` vs. `cloud_allowed`).</li><li><strong>De-Identification Controls:</strong> Built-in anonymization safeguards to strip personal information before executing steps.</li></ul><p>Using the automations interface, you can build workflows simulating tumor boards or heart teams where multiple experts research and synthesize information regarding medical cases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ukpB7JLIxQXIk8X_.png" /><figcaption>Tumorboard simulation example built in 2ndOpinion desktop app</figcaption></figure><h4>5. Specialized assistants &amp; guided chats</h4><p>You can also build assistants that can achieve specific tasks in the chat interface. you can start chat sessions using standardized forms for patient intake, literature reviews, or trial matching. You can easily define custom assistants with tailored system prompts, restricted toolsets, and form inputs. Again, you can force local models and/or de-identification steps.</p><h3>💻 Cross-Platform Availability</h3><p>2ndOpinion Desktop is available on:</p><ul><li>macOS (You may need to enable the app settings/privacy &amp; security)</li><li>Windows</li><li>Linux</li></ul><h4>🏁 Getting Started</h4><ol><li>Download the latest installer for your OS: <a href="https://2ndopin.io/desktop">https://2ndopin.io/desktop</a></li><li>Launch the app to automatically initialize your local databases.</li><li>Select your preferred model mode: download a local model via Ollama for 100% offline privacy, or enter your API key or use 2ndOpinion Cloud (default).</li><li>Start a research query!</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=339ab60230b2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to build your own tumor board simulation using AI]]></title>
            <link>https://briefs.aiadvances.org/how-to-build-your-own-tumor-board-simulation-using-ai-df0343435922?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/df0343435922</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[medicine]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 16:26:16 GMT</pubDate>
            <atom:updated>2026-09-04T13:12:57.570Z</atom:updated>
            <content:encoded><![CDATA[<p>Tumor boards are useful because difficult cancer cases are rarely just pathology problems, radiology problems, or surgical problems. The decision usually comes from putting all of those perspectives together. So, I wanted to see if we can simulate at least part of that process with an AI workflow.</p><p>The idea is not to build an “AI oncologist” that spits out treatment recommendations. Instead, the idea is to build something closer to the structure of an actual multidisciplinary discussion. This process essentially identifies what each specialist needs to clarify, lets those specialists independently address their part of the problem, and only then synthesizes the answers.</p><p>Before we start, <strong>here is </strong><a href="https://github.com/al2na/tumor_board_automation"><strong>the final workflow</strong></a><strong>.</strong> You can download and use it already if you don’t have the time to go through the rest of the post. In the link, you can also see the example case we used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NgCSRYq0uWCH5bov.png" /></figure><h4>The general idea</h4><p>The workflow starts with the available clinical information generally fed into tumor boards: 1) patient history, 2) pathology report, 3) radiology report, 4) a procedural or surgical report, and 5) the actual question we want the tumor board to answer.</p><p>The first task creates a short structured case summary. That summary and the original tumor-board question are then given to an “Oncologist” agent. This agent does not answer the question. Its task is to work out what the pathologist, surgeon, and radiologist should each be asked. Those questions then go to three separate specialist agents. Each specialist has access to the relevant case information and can search PubMed when external evidence is useful.</p><p>Finally, everything goes into a final synthesis node which acts as an oncology integrator. This task has to combine the specialist opinions, explicitly deal with disagreements and missing information, decide whether the case is actually ready for a decision, and only make a recommendation to the extent supported by the available evidence. Each of these tasks is represented as a separate node, so the behaviour of the individual “tumor board members” can be changed independently.</p><p><strong>Important caveat:</strong> This is obviously not an actual tumor board and should not be treated as one. This workflow is an experiment in structuring AI reasoning around multidisciplinary cases, not replacing clinical decision-making. All the steps rely on generative AI to generate task-specific text.</p><h4>Input nodes</h4><p>There are several inputs. In this case, all the inputs are simple text. The required inputs are:</p><ul><li>Patient history</li><li>Pathology report, including molecular biomarkers or other molecular tests</li><li>Radiology report</li><li>Diagnostic surgery/procedural report</li><li>The question we want the tumor board to answer</li></ul><p>The last one is particularly important. Without a focused question, it is very easy for an AI system to produce a giant generic discussion of the case rather than help with the actual clinical uncertainty. For example, the question could concern diagnosis, whether additional tissue is needed, resectability, disease extent, or what needs to happen before a treatment decision can be made.</p><h4>First node: Summarize the case</h4><p>The first agent acts as a clinical case summarizer. It takes the supplied reports and creates a short structured brief covering the clinical context, current problem, chronology, pathology, radiology, procedures, previous cancer treatment and relevant comorbidities or functional limitations.</p><p>There are some important constraints here. It is told not to diagnose, not to recommend treatment and not to resolve disagreements between reports.</p><p>The idea is to reduce a potentially long collection of reports into something the next agents can work with without quietly converting uncertainty into fact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GmvZ_vu0KDuRu0t7.png" /></figure><h4>Oncologist node</h4><p>Next task/node is the Lead Oncologist. This agent is also not allowed to solve the case. Instead, it receives the structured summary and the original tumor-board question and asks one or two focused questions of each specialist: Pathologist, surgeon and the radiologist.</p><p>The questions have to identify a specific uncertainty and ask the specialist for a direct conclusion, supporting case evidence, uncertainty and any additional information required before the conclusion can be relied upon.</p><p>So rather than asking the radiologist “What do you think?”, we want something closer to: what can the available imaging establish about disease distribution, and what remains uncertain in a way that could change management? Likewise, the pathologist should get the pathology-specific part of the problem rather than being asked to solve the entire oncology case.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2iKPFq4APUUKZ0f4.png" /></figure><h4>Pathologist node</h4><p>The Pathologist receives the focused pathology question plus the relevant clinical and pathology information.</p><p>The prompt is quite strict about separating four things:</p><ol><li>What is actually reported in the patient’s records.</li><li>The AI pathologist’s interpretation.</li><li>General knowledge supported by literature.</li><li>What remains uncertain.</li></ol><p>The agent is specifically instructed not to invent morphology, immunohistochemistry, molecular alterations, pathological stage, negative tests or specimen characteristics.</p><p>It can use PubMed when that helps with things such as differential diagnosis, interpretation of unusual immunophenotypes, molecular findings, classification criteria or additional tests that could separate competing diagnoses.</p><p>The output also forces the agent to explain exactly what uncertainty an additional test would resolve and whether resolving it could materially affect treatment.</p><p>That is important because otherwise “do more stains/molecular testing” can become the AI equivalent of ordering everything.</p><h4>Surgeon node</h4><p>The Surgeon gets a separate focused question from the oncologist. This agent looks at things like technical feasibility, the potential oncological role of surgery, diagnostic biopsy, possible surgical approaches, morbidity, functional consequences and what still needs to be established before a surgical decision can be made. Again, the scope is deliberately restricted. The surgeon is not allowed to establish pathology, choose systemic treatment, or produce the final multidisciplinary plan.</p><p>The surgeon node can also use PubMed to look at evidence around surgical indications, outcomes, neoadjuvant approaches, organ preservation and similar questions.</p><h4>Radiologist node</h4><p>The Radiologist performs the same kind of specialist analysis using the imaging report. The agent is explicitly told not to invent lesions, measurements, progression, anatomical relationships or metastatic disease. PubMed search is available for this task as well for questions around imaging modalities, response assessment, occult primary work-up, biopsy targeting and known limitations of different imaging approaches.</p><h4>Final synthesis node</h4><p>The outputs of specialist agents then go into the final synthesis node. This is the part where the simulated tumor board actually tries to answer the original question.</p><p>But before making a recommendation, it has to classify the case as: <strong>DECISION-READY, CONDITIONALLY DECISION-READY </strong>or <strong>NOT DECISION-READY</strong></p><p>A useful clinical decision-support system should be able to say “we cannot answer this yet” rather than treating production of a treatment plan as the definition of success.</p><p>The Integrator is also explicitly instructed not to silently resolve disagreements between specialists. If two specialist conclusions conflict, it has to state the conflict, explain why it matters and say what information would resolve it.</p><p>The final answer contains the integrated case formulation, specialist conclusions, reasoning, proposed sequence, conditions that need to be satisfied, alternatives and unresolved uncertainties.</p><h4>Privacy and clinical data</h4><p>One thing to note before using a workflow like this is patient data. The version of the workflow I am sharing is configured to allow cloud models, and the specialist nodes can send searches to PubMed. You can turn this behaviour off, so that only local models are used. If you want to keep using the cloud models on real patient cases, it would be prudent to de-identify it using the provided de-identification protocol, or you can do de-identification via other established means.</p><h4>How did the workflow do?</h4><p>We have tested this on a case report involving an <a href="https://academic.oup.com/oncolo/article/31/5/oyag132/8614627">adult neuroblastoma case</a>. The original publication is here. It classified the case as <strong>NOT DECISION-READY</strong> and asked for further evidence in the form of more molecular tests, including transcriptome or methylation tests. It mainly thinks that this is a neuroendocrine tumor with unknown primary, which is in the ballpark of adult neuroblastoma as it is rare for adults to have neuroblastoma. In the original paper, they confirmed the neuroblastoma after methylation tests; those tests were also not super definitive.</p><p>The problem with testing the workflow with such cases is that the case is either in the training set of large frontier LLMs or models can find it with a quick web search. So synthetic cases or cases that have no publication attached are the way to benchmark these systems. Nevertheless, the workflow described here can be adapted and tested for various cases. I encourage practitioners to have a go and contact me if they have questions.</p><h4>How to improve this further</h4><p>One can do a few or all of the following to improve this:</p><ul><li>Add a Medical Oncologist and Radiation Oncologist as independent specialist nodes.</li><li>Run the specialist agents more than once and add a consensus/synthesis step for areas where generative variability matters.</li><li>Add a dedicated evidence-quality agent that checks whether the cited PubMed papers actually support the propositions.</li><li>Add a contradiction detector before final synthesis to compare pathology, imaging, surgery and the original reports for disagreements.</li><li>Give the final Integrator an explicit mechanism for sending a question back to a specialist when something important remains unclear.</li><li>Add a patient-preferences/performance-status node when those factors are relevant to the clinical decision.</li><li>Make cancer specific guideline knowledgebases and use them in the evidence search steps. This is quite easy to do with the 2ndOpinion desktop app.</li><li>Compare it directly against simply giving the entire case to one LLM with one big tumor-board prompt. That is probably the most interesting experiment.</li><li>Make some nodes look at raw images from radiology or pathology with models like medGemma, 2ndOpinion can use such models in particular nodes.</li><li>Anything else you can think of, comment below or wherever you see this post.</li></ul><p>See you in the next post!!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df0343435922" width="1" height="1" alt=""><hr><p><a href="https://briefs.aiadvances.org/how-to-build-your-own-tumor-board-simulation-using-ai-df0343435922">How to build your own tumor board simulation using AI</a> was originally published in <a href="https://briefs.aiadvances.org/">AI Advances Briefs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Posting on substack from now on]]></title>
            <link>https://aakalin.medium.com/posting-on-substack-from-now-on-67ef3173048b?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/67ef3173048b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Mon, 19 Aug 2024 21:52:56 GMT</pubDate>
            <atom:updated>2026-08-27T15:39:07.522Z</atom:updated>
            <content:encoded><![CDATA[<h3>Posting on substack as well from now on</h3><p>I’ll be posting my content on Substack as well from now on. This platform offers a unique opportunity to connect more directly with readers.</p><p>I invite you to join the conversation there. You do not need a subscription to read content -&gt; <a href="https://substack.com/@al2na">https://substack.com/@al2na</a></p><p>Examples:</p><p><a href="https://biodatahacks.substack.com/p/how-to-build-your-own-tumor-board">https://biodatahacks.substack.com/p/how-to-build-your-own-tumor-board</a></p><p><a href="https://biodatahacks.substack.com/p/create-validated-case-briefs-from">Create validated case-briefs from clinical reports using LLMs</a></p><p><a href="https://biodatahacks.substack.com/">https://biodatahacks.substack.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67ef3173048b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI-assisted data analysis: How to make AI wrangle data for you]]></title>
            <link>https://aakalin.medium.com/how-to-make-ai-wrangle-data-for-you-f98dcb189128?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/f98dcb189128</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-wrangling]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Fri, 02 Aug 2024 08:38:55 GMT</pubDate>
            <atom:updated>2024-08-02T21:53:14.027Z</atom:updated>
            <content:encoded><![CDATA[<p>Most people doing data analysis enjoy the challenge of uncovering insights. However, We often find ourselves bogged down in the tedious task of cleaning and merging messy spreadsheets and tables instead. Fingers ache, eyes blur, and each small error feels like a mountain to climb.</p><p>I often ask if is there a quicker and less painful way to do these things, a way that would let us focus on meaningful analysis quicker…</p><h3>Why data wrangling is needed</h3><p>Data wrangling, also known as data munging, is a crucial step in the data analysis process. The primary goal is to transform raw data into a more useful and structured format for further analysis. Raw data often comes from various sources and is typically unstructured, incomplete, or inconsistent. This data can include errors, duplicates, and irrelevant information that hinder effective analysis. By cleaning and transforming the data, you ensure that it is accurate, complete, and consistent. This is essential for generating reliable insights from downstream analysis. It is also somewhat boring if you have many such tasks and you have to write similar code from scratch every time.</p><h3>What kind of tasks are included in data wrangling</h3><p>Data wrangling encompasses a variety of tasks aimed at transforming raw data into a usable format. One of the primary tasks is <strong>data cleaning</strong>, which involves identifying and correcting errors, removing duplicates, filtering data based on statistical criteria and dealing with missing values.</p><p>Another critical task is <strong>data transformation</strong>, which involves converting data into a format suitable for analysis. This can include normalizing data, aggregating information, and converting data types.</p><p><strong>Data integration</strong> is also a vital part of data wrangling, where data from different sources is combined into a single cohesive dataset. This task often requires matching records based on common fields and resolving discrepancies between different data sources.</p><p>Lastly, data wrangling could include <strong>data exporting</strong>. This might mean saving the cleaned and transformed data to a specific format or system for further analysis or sharing.</p><h3>How can AI help?</h3><p>Data wrangling can be tedious and repetitive. We can utilize LLM-driven frameworks models to quickly provide task adequate code and execution on your data. You simply need to type what you want to achieve on your dataset, rather than coding. We will be using the “<a href="https://github.com/BIMSBbioinfo/mergenstudio/">mergenstudio</a>” R package to demonstrate this concept. This package provides an LLM-driven framework on your Rstudio interface.</p><h3>Data filtering</h3><p>Let’s say you have some data that you want to filter based on some statistical measure. Writing the code for your analysis, performing the analysis and visualizing the results can be boring! Luckily, AI can help in both writing the analysis pipeline and visualization. Let’s look at an example of data filtering in gene expression data. We have some gene expression data from 6 different samples, across over 20,000 genes. We want to filter our data so that only the top 1000 most variable genes across samples are retained. With the help of <a href="https://github.com/BIMSBbioinfo/mergenstudio/">mergenstudio</a> this task becomes ridiculously easy! For this post, we assume you are already familiar with setting up your API information. If this is not the case, take a look at the GitHub help page (https://github.com/BIMSBbioinfo/mergenstudio). First, we select the directory where our data is at by clicking the button shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Sls2q7T35609NVqS" /></figure><p>After selecting the directory, we can start typing out our question. Important here is that we describe a bit what our datafile looks like, as to help the analysis along. After we give a brief description of our file, we ask can ask the LLM to perform the analysis for us. Here is the question we asked:</p><p>“ <em>I have a file called leukemiaExp.txt . It holds gene expression data for different samples. Samples are on the columns, and genes are on the rows. The first column holds gene names. Could you write code to filter the data so that only the top 1000 most variable genes across samples are retained? </em>“</p><p>So the answer we got back is the following:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FbiXYkzOsdXI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DbiXYkzOsdXI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FbiXYkzOsdXI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0d2d5290a2060db72e5d16ca48179e00/href">https://medium.com/media/0d2d5290a2060db72e5d16ca48179e00/href</a></iframe><p>As we can see, the LLM wrote code to help us perform the task at hand, and even wrote some optional code to save the filtered data to a file. So just by describing our file, and the type of data filtering we want to do, we get back a full data filtering pipeline which we can now run!</p><h3>Data transformation</h3><p>Data filtering is not the only type of wrangling that we can do. Other forms of data transformation can involve processes such as normalization, scaling, or converting data into a different format. One common example is transforming raw gene expression data from RNA sequencing (RNA-seq) into a normalized form that can be used for downstream analyses. Below is an example question in which we ask the LLMs help with data transformation and filtering afterwards:</p><p>“<em>I have a file called “RNAseq_coloncancer.txt” which holds RNA-seq data from different patients with different types of colon cancer. The rows hold different genes, and the columns represent different samples. The file is seperated by ‘,’ . I first want to log transform the RNA expression data. Since I have so many genes, I want to remove genes which have no real variance across samples. I want to retain the rows when there is 85% of distinct values out of the total number of samples. How do I do this?</em>”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iSiGLbKzbcvDxZm5" /></figure><p>As we can see, we can even execute the response, and see what the output of our analysis looks like!</p><h3>Merging tables</h3><p>Let’s face it. Dealing with data which is spread across multiple files can be pretty annoying. Luckily, AI can come to the rescue once more! Let’s take our raw gene expression data from RNA sequencing (RNA-seq) data. Oftentimes, we have some extra files holding additional information about the different samples or genes. So what if we want an easy way to merge these tables, but don’t know how to go about this? We can ask something like the following:</p><p>“<em>I have a file called “RNAseq_coloncancer.txt” which holds RNA-seq data from different patients with different types of colon cancer. The rows hold different genes, and the columns represent different samples. The file is seperated by ‘,’. I have another file called “patient_subtype.txt” which holds different information about phenotype in the column “subtype”. Could you merge these tables?”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fKIgqf6_7zB2oiXb" /></figure><h3>Basic plots after wrangling</h3><p>Data visualization is a crucial aspect of data analysis. Transforming your data sets into visual formats such as charts, graphs, and maps allows for easier interpretation, and helps to uncover patterns, trends, and insights that might not be apparent in raw data. Through compelling visuals, data becomes more accessible and engaging, driving informed actions and strategies. However, designing visually appealing plots can be laborious. <a href="https://github.com/BIMSBbioinfo/mergenstudio/">mergenstudio</a> can help with this! Imagine this case:</p><p>“<em>I have a file called “cancer_expr_set.txt” which holds the gene expression information of patients with different subtypes of blood cancer. Rows represent different genes, and columns represent the different patient samples. The file is seperated by “,”. The first 3 letters of the column names indicate which type of cancer the patient has. I want to plot a heatmap, and annotate this heatmap with the different cancer subtypes to see if they cluster together. How do I do this?”</em></p><p>We can get the desired heatmap by simply asking this question into the chat, and executing the response as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8q3YSECDig1r0AbN" /></figure><h3>Data output after filtering:</h3><p>Sometimes, after performing data wrangling, we would like to export some files holding the results from our analysis. Luckily, this is something we can also ask our LLM. Let’s take our first data wrangling example and modify it a bit:</p><p>“ <em>I have a file called leukemiaExp.txt . It holds gene expression data for different samples. Samples are on the columns, and genes are on the rows. The first column holds gene names. Could you write code to filter the data so that only the top 1000 most variable genes across samples are retained? After filtering I would like to save the information of the remaining genes in a txt file called leukemiaExp_filtered.txt.</em>”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lh0H3J20uCqNxbyQ" /></figure><p>If we now run this code and look into the selected directory, we will find the file called “<strong>leukemiaExp_filtered.txt</strong>” holding the results from the data filtering.</p><p>Try it and let us know how it went!!</p><p><strong><em>Disclaimer:</em></strong> Although <a href="https://github.com/BIMSBbioinfo/mergenstudio/">mergenstudio</a> works with almost any LLM API (OpenAI, Anthropic, Mistral, and also local LLM APIs via <a href="https://github.com/ollama/ollama/blob/main/docs/api.md">ollama</a>), for the examples above we used our API which is more performant than others for the data analysis tasks we care about. The performance you get from non-specialized LLM APIs might be different. If you want free access to our API, you can register your interest <a href="https://www.rolv.io/access-to-our-platform">here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f98dcb189128" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using artificial intelligence to find the best tumor models for cancer drug screening]]></title>
            <link>https://aakalin.medium.com/using-artificial-intelligence-to-find-the-best-tumor-models-for-cancer-drug-screening-a3c85e51b64f?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3c85e51b64f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[cancer]]></category>
            <category><![CDATA[medicine]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Thu, 09 Sep 2021 08:01:12 GMT</pubDate>
            <atom:updated>2021-09-22T13:43:58.214Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wCMSx72XRXNjbhRE" /><figcaption>Photo by <a href="https://unsplash.com/@_louisreed?utm_source=medium&amp;utm_medium=referral">Louis Reed</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The tumor models are necessary tools not only for the search for new drugs with antitumor activity but also for assessing their effectiveness. Optimizing the selection of the tumor models for drug screening is crucial for developing next-generation therapies for precision oncology. <strong><em>If the insights from these initial screens on tumor models were more transferable to primary tumors, more effective drugs could be developed</em></strong>. Needless to say, on top of the improved efficacy, a ton of money could be saved during preclinical and clinical trials by being able to design more precise trials and reducing the drug attrition rate.</p><p>Cell lines are widely used <em>in vitro</em> tumor models for drug discovery in basic cancer research. However, they are not perfect models of a tumor. Although cell lines are initially derived from patient samples, they might evolve towards different directions in cell culture. They also can not model the microenvironment of the tumor which affects drug response and overall survival. In addition, not all cell lines are equal and some are more similar to actual tumors when we consider their molecular features. Despite these difficulties, cell lines are one of the first steps to screen the efficacy of drugs that are in development. Since they are originally derived from a patient’s cancerous tissue, despite their divergence from the original tissue through cell culture over time, they still have some defects in their genome that would also occur in patients, therefore they are a good starting point for drug efficacy screening. Drugs are usually screened on a set of specific cell lines that have the target gene or defect of interest or broader screens that include a variety of genomic defects.</p><h3><strong>How can insights from cell lines be translated to </strong>patient-derived xenografts (<strong>PDX)?</strong></h3><p>PDXs (patient-derived xenografts) are tumor models, in which the tissue or cells from a patient’s tumor are implanted into an immunodeficient or humanized mouse. PDX can better recapitulate drug response in primary tumors than cell lines. They are generally the next step after a drug is proven to be a promising candidate on cell line models.</p><p>If you have a limited set of cell lines that you tested, which PDX models would reproduce the drug response you see in those cell lines? To answer this question, one must be able to measure the distance between the molecular features of cell lines and PDX somehow. The approach you take has to somehow transfer the information learned from cell lines to the PDX, and potentially also to the primary tumors. However, one important challenge for the transfer of such information between cell lines and the PDX models is that there are fundamental technical and biological differences between them. Yet, such differences can be irrelevant in terms of the models’ drug response outcomes. Therefore such irrelevant technical/biological differences should be sorted out or discounted.</p><p>A simple approach would be matching known genetic signatures that are associated with cancer between cell lines and PDX. However, that does not guarantee that the selected preclinical model will successfully represent the human disease or even the PDX. The drug response is more complicated than even the presence/absence of a mutation targeted by a drug. The presence or absence of a single mutation or a set of mutations does not guarantee transferability, we need to consider more dimensions, more molecular information when we make such decisions. However, when we consider more data types that define tumor models, it is also easy to make more mistakes when calculating the distance between cell lines and PDX.</p><h4><strong>Translating insights from cell lines into PDX: the wrong way</strong></h4><p>The figure below shows what happens when you don’t execute the information transfer properly and then calculate distances based on molecular features of cell lines and PDX. The scatterplots show the clustering of cell lines and PDX models based on the molecular features (gene expression, mutations, and copy number variation). As you can see, the PDX models and cell lines completely group into different clusters on the left-hand side plot. They also do not group by tumor tissue of origin (right-hand side plot). You will not be able to match your cell lines of interest to the right PDX models because of the primitive data integration techniques used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kXWehxtGjpGsMdKP" /><figcaption>Figure 1: Showing clustering based on molecular features of cell lines (CCLE: marked red on the left plot) and PDX (marked blue on the left plot). The right plot is identical to the left one just colored by the tumor origin of the samples.</figcaption></figure><h4><strong>Translating insights from cell lines into PDX: the right way</strong></h4><p>The data from cell lines and PDX models often do not have one-to-one correspondence. In addition, other batch effects such as differences in sequencing techniques can affect the comparability of cell lines to the PDX models. This creates insight transferability issues. The discordant datasets must be rectified to represent the same information during learning.</p><p><a href="http://arcas.ai/">Arcas technology</a> enables simultaneous data integration and rectification, where the machine learning method focuses on biological similarities between datasets and is not biased by technical differences. This way the data from cell lines and the PDX models can be integrated, thereby providing a way to match cell lines to their nearest neighboring PDX model and also to generate insights simultaneously from multiple data sets.</p><p>The figure below shows the results of proper integration of cell lines and the PDX models via the <a href="http://arcas.ai/">Arcas platform</a>. The cell lines and the PDX models are now mixed (left panel) and clustered according to the tumor of origin (right panel).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-qBFBBcBjpick2K6" /><figcaption>Figure 2: Showing clustering based on molecular features of cell lines (CCLE: marked red on the left plot) and PDX (marked blue on the left plot). Right plot is identical to the left one just colored by tumor origin of the samples.</figcaption></figure><p><strong>In summary,</strong> <a href="http://arcas.ai/">our technology</a> allows us to search and match cell lines to the PDX models, or search and match any other disease model to primary tumors. We also possess the expertise to deduce if the disease model has diverged to such an extent that it can no longer be a useful model for any primary tumor. This way, the insights obtained in pre-clinical phases of drug development can be projected to later stages more confidently resulting in better patient stratification and biomarker discovery, consequently cutting down costs and time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3c85e51b64f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to organize and run remote data science internships]]></title>
            <link>https://medium.com/data-science/how-to-organize-and-run-remote-data-science-internships-fb51ce21715?source=rss-c4165496d984------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb51ce21715</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[office-hours]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[editors-pick]]></category>
            <dc:creator><![CDATA[Altuna Akalin]]></dc:creator>
            <pubDate>Tue, 31 Aug 2021 09:07:28 GMT</pubDate>
            <atom:updated>2021-09-07T19:43:12.478Z</atom:updated>
            <content:encoded><![CDATA[<h4><a href="https://towardsdatascience.com/tagged/office-hours">Office Hours</a></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VDl6SOQezkDI779N" /><figcaption>Photo by <a href="https://unsplash.com/@jjying?utm_source=medium&amp;utm_medium=referral">JJ Ying</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>We are a <a href="http://bioinformatics.mdc-berlin.de/">team of scientists</a> working on computational methods for biological data analysis. One of our focus areas is the application of machine learning methods to genomics data. This year, we welcomed several interns for machine learning-related projects. Due to the pandemic, we had to run these internships remotely. However, remote internships are not bad at all and when set up correctly have even some advantages. The possibility of remote work provided us a much larger talent pool and in the end, we had the pleasure to work with very talented interns across 6 countries. The interns all had machine learning-related projects. Some were heavily focused on deep learning projects related to our <a href="http://arcas.ai/">Arcas AI platform</a>, and some focused on classical machine learning approaches on a variety of problems we are working on in the lab. Below is a summary of how we have structured these remote internship projects and what we learned in this process.</p><h3>1. Come up with well-structured projects</h3><p>The projects should be small and easy enough so that they can be completed in a matter of months. We chose projects that are related to some of the deep learning or classical machine learning projects in the lab. These projects usually but not exclusively rely on genomic data and have clinical applications. Most of the tasks revolved around modifying our existing approaches to improve the benchmark datasets we prepared.</p><p>Practically, we also provided notebooks for showing basic functionality for common data munging tasks that might be necessary. In addition, we set up goals and timelines to deliver those goals. We also made a conscious decision that all the code should be frequently committed to our GitHub repositories during the course of the projects and using a GitHub-based workflow for code development.</p><h3>2. Select the right people with the right skills</h3><p>Since the projects are well-defined, the skills needed to complete them are obvious. For deep learning projects, we required deep learning experience and for other classical machine learning projects, we required the relevant experience. It is hard to check if people have the bare minimum required experience or not, so whenever it was feasible we sent a data analysis challenge to the applicants. At the very least, we asked for a URL to GitHub or asked them to send us relevant coding projects to check their level of experience. In addition, we had a structured live interview process where we tried to gauge their level with questions such as: “Would you feel comfortable implementing a method described in a paper?” or asked them to describe their experience with a certain framework such as PyTorch or TensorFlow.</p><p>In all projects, the experience with a version control system, preferably git was a desired feature of the candidates. However, in some cases, we relaxed that requirement and that didn’t work very well. Some of the allocated time for those projects went into learning Git, which resulted in a loss of time that could have gone to improving the project deliverables. <em>Note to future self: Make sure people can use Git.</em></p><h3>3. Provide a pre-configured cloud compute environment</h3><p>This is a must for a smooth start. We provided cloud computing environments for the interns using AWS SageMaker but anything similar would do. The interns also had access to a“how to get started” document for the cloud environment set up and also the notebook providing a tutorial for how to use our deep learning-based method, <a href="https://github.com/BIMSBbioinfo/maui">Maui</a>, on genomic datasets.</p><p>There were occasions where we haven’t provided this and it took much longer for those projects to take off due to the time lost setting up a local compute environment and testing it. Creating a cloud computing environment is crucial to all future remote internship projects.</p><h3>4. Organize an introduction meeting</h3><p>If you are having multiple interns on related projects, it is always a good idea to introduce everyone to each other. This could be divided into project-based clusters of interns in case you have too many interns. However, groups of people under 10 would be easy to manage over zoom and everyone can introduce themselves and more importantly general project structures and brief onboarding can be done via these introductory meetings.</p><h3>5. Provide one contact point and supervision</h3><p>We organized the internships in a way that every intern has one direct supervisor that is chosen among the senior lab members. These direct supervisors were a point of day-to-day contact and provided much-needed support, especially in early phases. They also had regular meetings with the interns and monitored their progress and helped where needed.</p><h3>6. Keep communication easy and regular</h3><p>We opted in for slack to have an easy communication channel. Anything similar to slack works, we just picked it because we already use it. We created a special workspace for some of the larger internship projects. Having a dedicated chat app provides easy access to supervisors. We also held video calls where Slack was not quick enough or in cases, written communication would not be optimal. Sometimes getting on a call is quicker than writing issues on slack or elsewhere. In addition, many times interns collectively figured out how to solve problems by using slack, especially if it is a general problem due to the setup or the data you provided. You don’t have to provide all the guidance. If interns have a platform to communicate they can also guide each other.</p><h3>7. Communicate deliverables and milestones, keep the motivation high.</h3><p>We had goals and deadlines for those goals, and these were communicated to the interns. The deadlines were either decided between direct supervisors and interns for smaller weekly goals. Or they were related to joint meetings where overall progress must be presented, which we will touch upon next.</p><p>The deadlines etc. might create some pressure to reach the goals. However, they provide limited motivation. We saw that interns have lost motivation once in a while especially when the tasks seemed complex or when they couldn’t connect the tasks to larger goals. <strong><em>We found that the best remedy for that is to explain “why” repeatedly even when it is not directly asked</em></strong>. We need to remind interns why the work they do is important and the larger implications of that work. In our focus areas, this is not very difficult. Almost everything we do has to do with improved diagnostics or a better understanding of genome biology which might open up new areas or applications. These implications and how are they connect to an intern&#39;s tasks should be clearly and regularly communicated.</p><h3>8. Have joint meetings where progress and roadblocks are presented by all interns</h3><p>We held joint meetings for all the interns that work on related projects. This way interns get to know each other better and learn from each other during the meetings. We asked interns to present their progress and roadblocks so we can collectively address the roadblocks and celebrate the progress. The need for presenting results also created an additional motivation for interns to think about and summarize their work.</p><h3>9. Adjust your expectations. No matter what you do not everyone can be productive</h3><p>Things will not always work out. Some interns will just not be able to perform despite your best efforts to create structured projects and select the most promising interns. Although we were lucky that most interns hit their milestones, it is good to acknowledge that no amount of screening and planning can foresee unforeseeable events like medical emergencies or other health or family issues that will affect productivity.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb51ce21715" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/how-to-organize-and-run-remote-data-science-internships-fb51ce21715">How to organize and run remote data science internships</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>