Automating Repetitive Work Tasks

Few Lessons from Deploying and Using LLMs in Production Deploying LLMs can feel like hiring a hyperactive genius intern—they dazzle users while potentially draining your API budget. Here are some insights I’ve gathered: 1. “Cheap” is a Lie You Tell Yourself: Cloud costs per call may seem low, but the overall expense of an LLM-based system can skyrocket. Fixes: - Cache repetitive queries: Users ask the same thing at least 100x/day - Gatekeep: Use cheap classifiers (BERT) to filter “easy” requests. Let LLMs handle only the complex 10% and your current systems handle the remaining 90%. - Quantize your models: Shrink LLMs to run on cheaper hardware without massive accuracy drops - Asynchronously build your caches — Pre-generate common responses before they’re requested or gracefully fail the first time a query comes and cache for the next time. 2. Guard Against Model Hallucinations: Sometimes, models express answers with such confidence that distinguishing fact from fiction becomes challenging, even for human reviewers. Fixes: - Use RAG - Just a fancy way of saying to provide your model the knowledge it requires in the prompt itself by querying some database based on semantic matches with the query. - Guardrails: Validate outputs using regex or cross-encoders to establish a clear decision boundary between the query and the LLM’s response. 3. The best LLM is often a discriminative model: You don’t always need a full LLM. Consider knowledge distillation: use a large LLM to label your data and then train a smaller, discriminative model that performs similarly at a much lower cost. 4. It's not about the model, it is about the data on which it is trained: A smaller LLM might struggle with specialized domain data—that’s normal. Fine-tune your model on your specific data set by starting with parameter-efficient methods (like LoRA or Adapters) and using synthetic data generation to bootstrap training. 5. Prompts are the new Features: Prompts are the new features in your system. Version them, run A/B tests, and continuously refine using online experiments. Consider bandit algorithms to automatically promote the best-performing variants. What do you think? Have I missed anything? I’d love to hear your “I survived LLM prod” stories in the comments!

Instead of asking "what should I automate?" Focus on WHY you should automate and HOW it solves the data problem. Most data engineers automate the wrong things at the wrong time. Here's the framework I use after 8 years of building production systems: ✅ AUTOMATE WHEN: → Task runs daily/weekly → Human errors cause outages → Work blocks other priorities → Team growth = more manual work Examples: Reports, schema checks, alerts ❌ DON'T AUTOMATE WHEN: → Task happens quarterly → Requirements change weekly → Process isn't understood yet → Manual steps reveal insights My rule: If it’s done 3+ times, script it; 10+ times, automate it; fails 5+ times, redesign it. Automate what matters, when it matters—not everything! Here's how Airflow makes data automation ridiculously easy: 🎯 The Magic Triangle: → Scheduler: Triggers workflows on time → Executor: Distributes work to available workers → Workers: Actually run your Python code 💾 Smart State Management: → Metadata DB: Tracks every task run → Queue: Manages task priorities → Web UI: Visual monitoring & debugging 🔄 Why It Works: → Write Python DAGs once → Airflow handles the rest → Automatic retries & error handling → Parallel task execution → Visual dependency tracking Real Example: Instead of: ❌ Cron jobs that fail silently ❌ Manual dependency management ❌ No visibility into failures You get: ✅ Visual workflow monitoring ✅ Automatic failure notifications ✅ Smart task scheduling ✅ Easy debugging & restarting Image Credits: lakeFS The Bottom Line: Apache Airflow turns complex data workflows into manageable Python scripts. What's your biggest pipeline automation challenge? #data #engineering

Most coaches & consultants don’t have a time problem. They have a systems problem. AI doesn’t fix chaos. It scales whatever system you already have. Here are 5 AI tools that actually plug into your daily workflow (with real use-cases): 1. ChatGPT: Use it to think, not just write. Daily integration: Pre-call: Generate 5 sharp questions based on client background Post-call: Convert notes into insights and next steps Sales: Practice objection handling before discovery calls Example: “Here are my client notes → identify blind spots and suggest 3 tough questions for next session.” 2. Notion AI :Your second brain for client delivery. How to use: Create client dashboards with auto summaries Maintain SOPs for your programs Turn session transcripts into insights + next steps Example: Upload session notes → “Summarize key breakthroughs + assign action items” Your client gets clarity instantly. 3. Descript: Content creation without the headache. How to use: Edit podcasts/videos by editing text Remove filler words automatically Repurpose long-form content into shorts Example: Record a 20-min coaching insight → Cut it into 5 LinkedIn videos + 10 reels in under an hour. 4. Otter.ai.: Never miss what your client actually said. Daily integration: Record and transcribe coaching calls Highlight key patterns across sessions Build a repository of client insights over time Example: Spot recurring phrases like “I feel stuck” and use that language in your next session to go deeper. 5. Make: Where everything connects. Daily integration: Auto-send session summaries after calls Connect forms to CRM, email, and task managers Build end-to-end onboarding flows Example: Client fills a form, gets a calendar link, books a call, receives a prep doc, and you get a summary. All automated. Here’s the shift most people miss: Don’t ask, “Which AI tool should I use?” Ask, “Which part of my workflow is still manual?” That’s where AI fits. Because the goal isn’t to use more tools. It’s to free up more thinking time. What’s one task in your workflow you’d love to automate right now?

2025 saw a massive shift in how we perceive coding. It's 2026 now, and companies are still lagging behind. I used to think you needed developers to build products. Then I launched Searchable... And validated the entire idea with AI in 48 hours. At that level, I didn't need to know a single line of code. But if you're planning to replace real engineering work,  You'll need to create a proper plan of action. AI coding makes it easier than ever to build. But you still need to input clear ideas and know how it works. There are three levels of AI coding founders should understand: (See the visual for more details 👇) 1. Vibe Coding Level: Non-technical founders What it is: Turning rough ideas into working prototypes by describing what you want in plain English and letting AI handle the code. Business use case: → Validating startup ideas fast → Building landing pages, MVPs, internal tools → Testing demand before hiring engineers Tools to use: → Lovable - Product prototypes and signup flows → Bolt - Fast web app generation → Replit - Build and deploy without setup → Make - Connect tools and workflows 2. AI-Assisted Coding Level: Technical or semi-technical teams What it is: AI working alongside a human developer to speed up writing, debugging, and refactoring code. Business use case: → Building production-ready software faster → Improving developer output without growing headcount → Reducing bugs and repetitive work Tools to use: → Cursor - AI-first code editor → GitHub Copilot - Inline code assistance → Continue - Open-source AI coding assistant → Google Antigravity - Context aware completions 3. Agentic Coding Level: Advanced team and operators What it is: AI agents that can plan, write, test, and refine entire chunks of software from a single objective. Business use case: → Large feature builds → Legacy code refactors → Automating repetitive engineering tasks → Spinning up internal systems fast Tools to use: → Claude Code - Agent-driven deployment → OpenAI Codex - Autonomous coding tasks → Devin - Full software agent → Gemini CLI - Command-line agent workflows These tools let you validate first and hire second… Yet another way AI allows founders to move faster than ever before. If you’re building right now, this is leverage you can’t ignore. Are you familiar with AI coding? How are you using it?  Drop a comment below with your process.  At Searchable, we're using AI to build an autonomous SEO and AEO growth engine. It analyses, fixes, and scales websites to drive customers automatically. If you're a founder who wants to stay visible when people search with ChatGPT, Perplexity, or Google AI... This is built for you. Learn more and get started with a 14-day free trial here:  https://lnkd.in/epgXyFmi ♻️ Repost to share this breakdown with founders in your network.  And follow Chris Donnelly for more on building smarter. 

Pick the tool based on the task, not the logo. AI coding tools are no longer just autocomplete assistants. They are becoming different layers of the developer workflow. Claude Code, OpenAI Codex, Cursor, and GitHub Copilot all help developers move faster, but each one is strongest in a different situation. Claude Code fits well when you need deep repo work, terminal-native execution, multi-file changes, refactoring, test runs, and autonomous codebase updates. OpenAI Codex is useful when you want to delegate coding tasks, run work in the cloud, parallelize fixes, and let background agents return results for review. Cursor works best when you live inside the IDE and want fast daily coding, inline edits, codebase chat, agent changes, and quick iteration. GitHub Copilot is strong for teams already working inside GitHub, especially for autocomplete, PR support, repo chat, reviews, and enterprise adoption. The mistake is asking: “Which one is the best?” The better question is: “What job do I need this tool to perform?” For active development → Cursor For repo-wide refactors → Claude Code For async delegated tasks → OpenAI Codex For GitHub-native team workflows → GitHub Copilot The future is not one AI coding tool replacing every other tool. It is a stack. One tool for writing. One tool for refactoring. One tool for delegation. One tool for review and shipping. AI coding has moved from suggesting lines of code to helping developers plan, edit, test, review, and ship across real codebases. The real advantage will come from knowing which tool to use at which stage of the workflow. Which AI coding tool fits your current development workflow best?

Many people I talk to have heard of coding agents and are interested in using them, but don't know: 1. What current coding agents can do 2. How users can prompt agents effectively To help out with this, I wrote a blog on 8 use cases for coding agents, with example prompts: https://lnkd.in/gdizhYMX The first two use cases are familiar ones, (1) fixing bugs and (2) adding features. This is the common "resolve github issue" usecase tested in benchmarks like SWE-Bench. One nice thing about good agents is that they can implement a change and test it. Here's an example prompt. Another thing people are using agents for a bit is (3) creating new apps from scratch. Here's are two examples for a frontend app and email sending script that I successfully created with OpenHands. In this case I prompt about some design decisions, like what framework to use. The next three are actually my favorites: (4) fixing failing CI tests, (5) fixing merge conflicts, and (6) writing docs. These tasks every developer needs to do but noone wants to do. It's been a huge boost to be able to ask the agent to resolve these issues for me; it generally works well! Coding agents can also (7) help with deployments by spinning up cloud resources. Obviously you need to carefully supervise the agent to make sure that it doesn't break anything, but with careful credentialing and review of infrastructure as code it can make deployment much easier! Finally, I have been using coding agents a lot for (8) data analysis tasks. I asked OpenHands to create a script to monitor commit activity on our repo, and the resulting graph is in the blog. One final easter egg, I actually asked OpenHands to make the header figure at the top of this thread too! I asked it: * Download appropriate from fonts-awesome * Arrange them 2 rows and 4 columns center-justified with the text * Make them rainbow colored * Write to png If any of these use cases sound interesting, I'd encourage you to read the blog and try out OpenHands, a general software development agent that can help with these tasks: * Download now: https://lnkd.in/g4VhSi9a * Sign up for the web app: https://lnkd.in/gJ-_SFv2

Few-shot Text Classification predicts the label of a given text after training with just a handful of labeled data. It's a powerful technique for overcoming real-world situations with scarce labeled data. SetFit is a fast, accurate few-shot NLP classification model perfect for intent detection in GenAI chatbots. In the pre-ChatGPT era, Intent Detection was an essential aspect of chatbots like Dialogflow. Chatbots would only respond to intents or topics that the developers explicitly programmed, ensuring they would stick closely to their intended use and prevent prompt injections. OpenAI's ChatGPT changed that with its incredible reasoning abilities, which allowed an LLM to decide how to answer users' questions on various topics without explicitly programming a flow for handling each topic. You just "prompt" the LLM on which topics to respond to and which to decline and let the LLM decide. However, numerous examples in the post-ChatGPT era have repeatedly shown how finicky a pure "prompt" based approach is. In my journey working with LLMs over the past year+, one of the most reliable methods I've found to restrict LLMs to a desired domain is to follow a 2-step approach that I've spoken about in the past: https://lnkd.in/g6cvAW-T 1. Preprocessing guardrail: An LLM call and heuristical rules to decide if the user's input is from an allowed topic. 2. LLM call: The chatbot logic, such as Retrieval Augmented Generation. The downside of this approach is the significant latency added by the additional LLM call in step 1. The solution is simple: replace the LLM call with a lightweight model that detects if the user's input is from an allowed topic. In other words, good old Intent Detection! With SetFit, you can build a highly accurate multi-label text classifier with as few as 10-15 examples per topic, making it an excellent choice for label-scarce intent detection problems. Following the documentation from the links below, I could train a SetFit model in seconds and have an inference time of <50ms on the CPU! If you're using an LLM as a few- or zero-shot classifier, I recommend checking out SetFit instead! 📝 SetFit Paper: https://lnkd.in/gy88XD3b 🌟 SetFit Github: https://lnkd.in/gC8br-EJ 🤗 SetFit Few Shot Learning Blog on Huggingface: https://lnkd.in/gaab_tvJ 🤗 SetFit Multi-Label Classification: https://lnkd.in/gz9mw4ey 🗣️ Intents in DialogFlow: https://lnkd.in/ggNbzxH6 Follow me for more tips on building successful ML and LLM products! Medium: https://lnkd.in/g2jAJn5 X: https://lnkd.in/g_JbKEkM #generativeai #llm #nlp #artificialintelligence #mlops #llmops

Managing prompts is messy. You tweak a word here, remove a shot there, rerun the pipeline — and still don’t get the results you expect. It’s frustrating. And not scalable. That’s where 𝗔𝗱𝗮𝗹𝗙𝗹𝗼𝘄 comes in — a PyTorch-style library designed to 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲 𝗮𝗻𝗱 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗲 your prompt workflows for RAG and AI agents. You define a tunable prompt, pick your LLM, set a reward function, and let AdalFlow handle the scoring and tuning — across multiple examples. No gradients. No guesswork. Just smarter, faster prompt iteration. Why it matters: • Supports black-box reward functions • Works with OpenAI, Groq, and others • Auto-optimizes few-shot and zero-shot prompts • Drop-in ready for your current LLM stack • Prompt engineering shouldn’t feel like debugging JavaScript in production. Link to the GitHub repo in the comments.

When optimizing prompt for #LLM, what if we don't have a score to evaluate performance of every prompt? The work of Xiaoqiang Lin, Zhongxiang Dai, Arun Verma, See-Kiong Ng, Patrick Jaillet, and Bryan Kian Hsiang Low achieves prompt optimization with human feedback (APOHF) only. APOHF uses a dueling bandit strategy to choose a pair of prompts in every iteration to generate their responses which are then shown to the user for preference feedback. In practice, the user only needs to give an initial task description & then a series of preference feedback. We use APOHF to optimize the prompt for #DALLE3 while only asking user for preference feedback between pairs of generated images. APOHF can efficiently produce an image which aligns well with the user's preference (compare ground truth image and image at iteration 10) while only requiring a small number of human feedback instances. APOHF can be adapted to solve response optimization with human feedback. For every received prompt, let the LLM generate many responses. Then, adapt APOHF to choose a pair of responses to query the user for preference feedback. As shown in the table, the response discovered by APOHF aligns well with human preferences as it is well organized (via a numbered list) and detailed. https://lnkd.in/g_Gv5Wse #LLMs #RLHF

Someone spent months reverse-engineering every Claude Code feature into one free guide. 11K stars. 690 forks. Here's what most Claude Code users don't know exists: → /self-assessment runs directly inside Claude Code - get a personalized learning roadmap based on what you already know → /lesson-quiz [topic] after each module pinpoints exactly what you missed → Hooks trigger automatic actions before and after every Claude response - no manual intervention → Checkpoints let you rewind your entire session to any previous state → Skills teach Claude your team's architecture patterns once - it follows them forever → Subagents run specialized tasks in parallel - code review, security audits, documentation in one pipeline → Plugins bundle everything into one installable unit your whole team can use → EPUB generation built in - one script, entire guide as an offline ebook What you can build when you combine them: → Automated code review: Slash Commands + Subagents + Memory + MCP → CI/CD automation: CLI + Hooks + Background Tasks → Security audits: Subagents + Skills + Hooks in read-only mode The questions this repo answers: → When should you use a slash command vs a subagent vs a skill? → How do you wire MCP into an automated pipeline? → What does a production-ready Claude Code workflow actually look like? GitHub Repo: https://lnkd.in/dsGA2egQ