Ask HN: What have you built with LLMs?

Hacker News Discussion ↗

Developers are using large language models to power a huge range of projects, from Chrome extensions, games and voice bots to code assistants, RAG-based knowledge tools, news and video summarizers, and niche productivity apps. Many experiments focus on augmenting existing workflows — automating note-taking, email and hiring, translating media, personalizing learning, or making local data searchable — often with small stacks built on OpenAI, open‑source models, and vector databases. Recurring themes include the ease of prototyping, challenges around latency, cost and hallucinations, and a sense that many LLM products are quickly commoditized unless tied to specific domains or unique data.

Range of applications

  • Developer & data tools

    • “Chat with your data” for SQL/CSV/Jira/Salesforce; natural-language → SQL CLIs.
    • RAG frameworks, vector DB services, and GitHub/GitLab repo Q&A assistants.
    • Prompt management systems, agent hosts, and tools to auto-generate diagrams, code modifications, or PDFs.
    • Robot task planners and browser-controlling agents (e.g., for generating tests).

  • Productivity & business workflows

    • Email/Gmail assistants, calendar companions, hiring screeners, sales/CRM helpers, order and shipment trackers.
    • Tools to summarize or analyze calls, meetings, help docs, and drilling/geology reports.
    • Internal firm assistants for law and other professional services; credentialing/licensing cost summarizers.

  • Content summarization & transcription

    • YouTube / podcast / news / HN / RSS summarizers.
    • Subtitle/SRT translators (including slang-heavy languages) and Estonian ASR.
    • Whisper-based pipelines for podcasts, books, and live meeting assistants.

  • Learning, training & education

    • Language-learning apps (spaced repetition, chat partners, pronunciation, grammar correction).
    • Mock coding interviews, sales-call simulators, tabletop-GM helpers, homeschool tutors, resume chatbots.
    • Tools to auto-generate slides, videos, and course structures from text.

  • Creative, games & entertainment

    • Text adventure games, Turing-test chat games, murder mysteries, crafting games, card/skin generators.
    • Story and children’s-book creators, absurd news/personality generators, autonomous space-opera TV shows.

  • Domain-specific tools

    • Finance: earnings-call summarization, news briefings, job-search engines.
    • E‑commerce: image-based product-description generators, receipt scanners, caloric analyzers, product-photo describers.
    • Accessibility: alt-text generators and AAC apps for autistic users.

Technical stacks & deployment

  • Frequent stacks: Next.js/React/TypeScript, Python (FastAPI/Flask), Deno, SvelteKit, Flutter, Elixir, Go, Rust, Kotlin.
  • LLMs: OpenAI GPT-3.5/4(+Vision), local models via Ollama/llama.cpp/vLLM, Mistral/Mixtral, Whisper, custom STT/TTS.
  • Infra: Vercel, Cloud Run, Deno Deploy, Render, bare metal, Docker/Kubernetes; vector backends like pgvector, ChromaDB, Supabase, FAISS.

Patterns, limitations & lessons

  • LLMs praised for rapid prototyping, glue code, and making previously infeasible UX (e.g., parsing arbitrary emails, receipts).
  • Cost and latency are recurring pain points; many move to local or cheaper models and heavy caching.
  • Hallucinations, fragile JSON/output schemas, content filters, and context limits require retries, multi-pass pipelines, and evaluation loops.
  • RAG quality depends heavily on chunking, retrieval, and domain-tuned prompts; several tools focus solely on benchmarking/optimizing this.
  • Many note how easily LLM products are commoditized; differentiation often comes from domain focus, UX, or integrations rather than model choice.
  • LLMs seen as especially valuable for “risk-free practice” (sales, interviews, language) and for taming unstructured or tedious text.