12 GenAI Projects That Prove You Can Build
LinkedIn post by Aryan Raj (Founder @ Planck Labs, “Master AI before it masters you”) — view post
Premise: “Most people stay stuck watching tutorials.” These 12 projects each signal a specific production skill to employers:
| # | Project | What it is | What it shows |
|---|---|---|---|
| 1 | Production API Wrapper | FastAPI proxy with error handling, rate limiting, webhook support | production APIs, not just scripts |
| 2 | Token Cost Estimator | Calculates spend before generation; usage per endpoint; budget-threshold alerts | economics, not just capabilities |
| 3 | Validated JSON Agent | Pydantic schemas, retry on parse errors, logged validation failures | preventing parse errors in prod |
| 4 | Cited RAG Bot | PDF Q&A with hybrid search + reranking + page-specific citations | grounding outputs, cutting hallucination |
| 5 | Human-in-the-Loop Workflow | Research agent pauses for approval on expensive actions; full audit trail | safe multi-step orchestration |
| 6 | Streaming Copilot UI | Real-time token streaming, graceful latency degradation, optimistic UI | shipping user-facing products |
| 7 | Automated Eval Harness | Scores accuracy on 50 golden test cases; catches regressions pre-deploy | measuring quality, not guessing |
| 8 | Local Inference Server | vLLM deployment with KV-cache optimization and quantization; zero API costs | performance under the API layer |
| 9 | Traced Deployment Pipeline | Dockerized agent + distributed tracing, cost dashboards, failure alerting | debugging prod issues in minutes |
| 10 | Security Guardrail Middleware | Blocks prompt injection, redacts PII, enforces I/O filtering policies | enterprise-safe systems |
| 11 | Public Architecture Repo | Architecture docs, demo video, benchmark reports | communicating technical decisions |
| 12 | Portfolio Site | Links all 11 with value props + live demos | ready to work, not just learn |
Comment-worthy additions: commenters endorsed 7 (evals) and 9 (tracing) as the ones that matter most once a real system exists (“evals and observability… make sense once you have something running that you’re afraid to break”); another recommended evaluating a validated-JSON agent first since schema enforcement is where pipelines actually break. One reply pointed out evals were already covered by #7.