What a 24-hour AI PoC must prove before you build
A technical proof of concept should never be a beauty contest for a user interface. Its job is to verify in principle that a working AI pipeline can solve the core or narrow technical bottleneck of your idea under realistic latency and unit-cost bounds, so leadership can make an evidence-based Go, Pivot, or Kill decision in under a week.
Four numbers every AI PoC must produce
- Task accuracy and groundedness on messy cases. Not five cherry-picked happy prompts, but 20–50 domain examples including ambiguous queries, missing documents, and noisy inputs.
- Refusal both ways. How often the pipeline answers when it should refuse (policy or missing evidence) and how often it refuses valid user requests.
- Latency distribution (p50 and p95). End-to-end response time under realistic context lengths and tool-call chains.
- Unit cost per successful task. Total token, embedding, and reranker spend per completed workflow, including retries.
How Arviser runs a 24-hour PoC Spike (from €2,400)
- Hours 1–4: Lock the target pipeline slice, measurable pass marks, and a 20–50 case test set in writing.
- Hours 5–16: Senior engineer (10+ years experience) builds an isolated CLI or notebook harness comparing 2–3 candidate models or retrieval designs.
- Hours 17–21: Automated evaluation run in Promptfoo or DeepEval across all candidate setups.
- Hours 22–24: Written Go, Pivot, or Kill memo, unit-economics projection, full repository handoff, and founder readout.
Risk-shared 50% guarantee and fee credit
- When feasibility is confirmed (Go): Even if only a focused slice of the task is needed to prove the pipeline works, full payment (€2,400) transfers 100% exclusive IP ownership under NDA, and 100% of the fee credits toward an 80-hour Production AI MVP, Pilot-to-Production, or LLM Quality Audit within 60 days (or 50% toward a 40-hour Working Prototype Sprint).
- When the hypothesis does not hold (Kill): Under our Risk-Shared Open-Research Track, the second 50% payment is waived (you pay €1,200 total). You keep a non-exclusive license to the code and full benchmark report, while Arviser retains harness rights and may publish a sanitized, synthetic-data benchmark note.
Next: 24h Technical PoC Spike · Risk-shared PoC protocol · Startup Program · Request scope