Real AI engineering builds with full build environments. Pick one, build it, ship it.
12 ideas
RAG-001 · RAG & Knowledge
Production retrieval systems fail silently. Build an orchestrator that detects poor retrieval quality and reroutes queries using fallback strategies — before the user sees a bad answer.
AGT-002 · Agents
Most agent frameworks can't remember what they did two steps ago. Build a stateful agent runtime with episodic memory, reflection loops, and rollback capability.
LLM-003 · LLM Infra
Serve LLMs efficiently at scale. Implement dynamic KV cache management, request batching, and latency-aware routing across model replicas.
DAT-004 · Data & Features
ML models degrade in production without warning. Build a streaming drift detector that compares live feature distributions against training baselines and alerts on deviation.
SAF-005 · Safety
You can't deploy what you can't evaluate. Build a modular safety harness that runs adversarial probes, jailbreak tests, and alignment checks across any LLM endpoint.
VEC-006 · Vector & Search
ANN search is at the core of every RAG system. Implement HNSW from scratch, expose a REST API, and benchmark it against production vector databases.
RAG-007 · RAG & Knowledge
Single-retrieval RAG breaks on complex questions. Build a pipeline that decomposes multi-hop queries, retrieves across sub-questions, and synthesizes a grounded answer.
AGT-008 · Agents
LLM tool-calling is brittle without validation. Build an orchestrator that wraps tool calls with schema enforcement, retry logic, and structured error recovery.
LLM-009 · LLM Infra
Most LLM apps re-compute answers to semantically identical prompts. Build a caching layer that stores completions by semantic similarity, not exact hash.
DAT-010 · Data & Features
Training/serving skew kills production models. Build a lightweight feature store with point-in-time correctness, online serving, and offline batch access.
SAF-011 · Safety
User data leaks through LLM prompts constantly. Build a preprocessing pipeline that detects and redacts PII before it reaches any model endpoint.
VEC-012 · Vector & Search
Dense retrieval misses keyword-critical queries. Build a hybrid search system that blends BM25 and dense embeddings, then re-ranks with a cross-encoder.
More ideas added every month. Paid members get instant access.