Five Weekend AI Projects That Teach Production Engineering, Not Prompt Tricks

Every project on this list attacks an operational problem that companies running LLMs at scale actually pay to solve: cost, latency, regressions, failure tracing, and stale docs. Bashi, a software engineer who posts as @bashifuirkashi, walks through five buildable weekend projects and claims they will teach you more than any course. The video is a screen recording of a written guide, so the frames show full tech stacks and step-by-step build phases for each project while he narrates the one-line pitch. The through line is that each project treats AI engineering as a business problem with measurable ROI, which is the framing hiring managers reward.

Project 1: LLM Cost Autopilot

The spoken pitch is a routing layer that sits in front of multiple LLM providers, analyzes the complexity of each incoming query, and routes it to the cheapest model that can still answer with quality. The guide page shown on screen adds the business case: companies running LLMs at scale overspend on over-provisioned model calls, and building a cost optimizer signals you understand AI engineering as a business problem. The visible tech stack is Python 3.11+, OpenAI, Anthropic, and Ollama as providers, FastAPI as the router, scikit-learn or a small fine-tuned model as the complexity classifier, custom scoring plus LLM-as-judge for eval, SQLite with structured JSON logs, Streamlit or Grafana for dashboards, and Docker Compose. Phase 1 in the guide starts with a model registry holding cost per input and output token and a quality tier for each model.

Project 2: Semantic Caching Layer for LLM APIs

This is middleware that detects semantically similar requests, stores the answers, and serves cache hits instantly to cut latency and API cost. The guide page on screen quantifies the goal as near-zero latency on hits and a 30 to 60 percent reduction in API costs on typical workloads. The stack shown uses OpenAI text-embedding-3-small for embeddings, Redis with RedisVL or Qdrant as the vector store, FastAPI as a drop-in proxy, a custom TTL plus similarity threshold cache policy, and Prometheus with Grafana for hit-rate monitoring. The build guide is specific about the core mechanic: embed every incoming prompt, run a nearest-neighbor lookup, and treat cosine similarity above a threshold (starting at 0.95) as a cache hit.

Project 3: Model Regression Detection System

Bashi describes this as a CI/CD-style pipeline that continuously tests any LLM-powered feature against a golden dataset, then alerts the team on Slack when a system prompt or model change causes a quality regression. The on-screen rationale is blunt: every AI team ships prompt changes blind, and this project proves you think about what happens after deployment. The stack shown is the OpenAI API, a custom eval framework with RAGAS or DeepEval, SQLite plus JSON files for storage, Slack webhooks for alerting, GitHub Actions for scheduling on every PR, and Streamlit or an HTML report for diff views. Phase 1 has you build a small email classifier as the feature under test and version your prompts as files in a prompts directory.

Project 4: Failure Forensics Tool for AI Pipelines

This one is an observability layer for multi-step AI pipelines. It traces every intermediate step, identifies exactly where failures originate when the final output is bad, and flags failures for evaluation. The guide frames it as building a mini LangSmith or Braintrust, and calls the ability to articulate why observability matters for AI systems a senior-level signal. The stack on screen uses a custom chain or LangChain for the pipeline, OpenTelemetry with custom spans for tracing, SQLite plus JSON trace files for storage, a React frontend or Streamlit for visualization, and a simple REST API for the feedback loop. The starter exercise is a four-step document pipeline covering intake, extraction, classification, and summary.

Project 5: Self-Healing Technical Docs

The final project is a GitHub Action that monitors a codebase, detects when code changes make documentation inaccurate, identifies the stale sections, and either auto-generates a PR with corrected docs or flags discrepancies for human review. The guide argues this project stands out because it lives inside a CI/CD pipeline rather than a Streamlit demo, and it exercises the full AI engineering stack: embeddings, retrieval, LLM generation, and production deployment. The stack shown is Python or TypeScript, OpenAI text-embedding-3-small, file-based ChromaDB so no server is needed, GPT-4o or Claude Sonnet for code understanding, and PyGithub with git diff for PR creation.

The Video Is a Lead Magnet for a Paid Guide

The projects come from a written guide titled "15 AI Engineering Projects That Actually Land Jobs," shown on screen with the branding AiEngineerAccelerator and BASWE, and the URL skool.com/baswe-ai. That URL resolves to the BASWE.Ai Engineer community on Skool, a paid group ($89 per month at the time of checking) run by Bashiri Smith for engineers moving into AI roles. The video ends with a comment-gate: comment "five" and he sends the links. The five projects in the video map to guide entries numbered 1, 2, 3, 4, and 7, so the video is a curated sample of the 15.

Key Takeaways

• All five projects target LLM operations problems with measurable dollar impact: routing cost, cache hit rates, quality regressions, failure location, and doc drift.
• The guide repeatedly frames each build as interview signal, arguing that thinking about cost, evals, and observability separates senior candidates from junior ones.
• The stacks are deliberately lightweight and reproducible: Python, FastAPI, SQLite, Docker, and GitHub Actions appear in nearly every project, with no managed infrastructure required.
• Recurring evaluation patterns show up across projects: golden datasets, LLM-as-judge scoring, and cosine similarity thresholds around 0.95 for semantic matching.
• The video samples 5 of 15 projects from a paid guide tied to the creator's Skool community, so treat the project list as solid and the "land jobs" claim as marketing.

Resources

• BASWE.Ai Engineer community on Skool - the community behind the "15 AI Engineering Projects That Actually Land Jobs" guide shown in the video (verified; paid membership).
• Tools shown on screen in the guide's tech stack tables, all standard open source or commercial products: FastAPI, scikit-learn, SQLite, Streamlit, Grafana, Redis/RedisVL, Qdrant, Prometheus, RAGAS, DeepEval, LangChain, OpenTelemetry, ChromaDB, PyGithub, Docker, GitHub Actions. No individual links verified for these.