Our AI features cost real money per use. Do you bundle into existing plans, charge per use, or create an AI add-on tier?
Shipping LLM features to production, not just demos
146 members 9 postsTechnology
For engineers and data folks building with LLMs: RAG, evals, agents, fine-tuning, cost and latency. Show what you shipped, what broke and what you learned.
Our AI features cost real money per use. Do you bundle into existing plans, charge per use, or create an AI add-on tier?
Arjun Mehta
Founder @ Shiftbase Labs (B2B SaaS for shift scheduling) | Ex-PM | Building in public
Not the fancy one. It's the 'summarise this week's shift swaps and absences' email for managers. Boring, useful, used daily by 80% of managers.
Priya Banerjee
Data Scientist | NLP, forecasting, Python | Ex-researcher | Women in Tech mentor
We replaced an LLM call for intent classification with a fine-tuned small model. Latency 1.2s → 40ms, cost basically zero, accuracy the same. Not everything needs a frontier model.
Aarav Sharma
Senior Backend Engineer @ Paylane | Java, Spring Boot, Kafka | Building payment rails that don't wake you up at 3am
Our docs bot sometimes invents config flags that don't exist. We already tell it to only use the context. What else works?
Ishaan Reddy
ML Engineer | LLMs, RAG, Python | Shipping AI features that survive production
Just curious what people are actually running, not what they're experimenting with.
Every AI feature spec I review now has to answer: 1. How does the user know this was AI-generated? 2. How do they correct it? 3. What's the cost of a wrong answer, and do we gate on confidence? 4. What do we log so we can debug it later?
Priya Banerjee
Data Scientist | NLP, forecasting, Python | Ex-researcher | Women in Tech mentor
Working on call-transcript summaries. We don't have gold summaries for most data. Currently using an LLM judge with a rubric (coverage, faithfulness, conciseness), but I'm not sure how much to trust it. What's worked for others?
Ishaan Reddy
ML Engineer | LLMs, RAG, Python | Shipping AI features that survive production
We built a golden set of 300 real support questions with expected answers, then measured each change: • Better chunking (by heading, ~400 tokens, 15% overlap): +6 • Hybrid search (BM25 + embeddings): +8 • Cross-encoder reranking of top 30 → top 5: +7 • Prompting the model to cite chunk IDs and refuse if none apply:
Ishaan Reddy
ML Engineer | LLMs, RAG, Python | Shipping AI features that survive production
This community is for people shipping LLM features to real users. Demos are fun, but let's talk about what happens after: evals, latency, cost, failures, user trust. Introduce yourself: what are you building and what's the hardest part right now?