Avenli

The problem

Complex decisions scatter goals, constraints, research, trade-offs and next steps across tabs. A one-off chat answer can mix supported facts with assumptions and leaves little durable record of how a recommendation was reached. Avenli helps people facing a time-bounded personal or work decision understand the evidence, remaining uncertainty and next actions in one place.

What Avenli does

The user describes an outcome, deadline, constraints and unknowns. Eight bounded stages turn that input into a structured goal, research plan, source-linked evidence, checked claims, compared options, an adversarial critique, a conservative recommendation and proposed tasks. The workflow saves checkpoints, survives interrupted requests, can pause for a human clarification and resumes the same run. Results remain inspectable; tasks can be tracked without automatically executing them.

The authenticated workspace supports saved problems, due dates, in-app reminders, focused follow-up analyses and bounded daily condition monitoring. Optional Gmail and Calendar context can be enabled separately. A Calendar event is the only implemented external write: Avenli shows the exact proposed event and creates it only after separate user approval, with idempotency and an audit trail. Agents themselves have no executable tools.

Why Avenli instead of a one-off AI chat

General assistants such as ChatGPT and Gemini are useful for conversation. Avenli is built around the life of a problem: a saved workspace shows real completed stages, pauses for essential missing details, keeps source-linked claims and uncertainty inspectable, and turns the recommendation into trackable proposed tasks. It gives the user a clear next step after research and requires separate approval before a real-world action. This is a workflow distinction, not a claim that every answer is more accurate.

How Nebius, NVIDIA and Tavily power it

Avenli calls NVIDIA nvidia/nemotron-3-super-120b-a12b through Nebius Token Factory for all eight role-specific stages: Intake, Planner, Researcher, Verifier, Options, Critic, Decision and Tasks. We chose Super for structured, multi-stage reasoning that remains practical for repeated calls. Every stage uses a strict output contract and cross-stage reference checks. The app can reject or narrowly repair invalid read-only output; it never lets the model invent executable tool calls.

Tavily makes functional runtime searches for up to three priority research questions. The server canonicalizes and deduplicates results, caps the evidence set and treats retrieved text as untrusted. Claims may cite only application-assigned source IDs. Source quality, freshness, contradictions and incomplete coverage can lower confidence or make the recommendation abstain. Nebius provides hosted inference through an OpenAI-compatible API without requiring us to operate GPU infrastructure.

How we built and checked it

Codex helped translate the product plan into the Next.js/Supabase architecture, implement migrations and row-level security, build the checkpointed orchestrator, integrate providers and Google OAuth, refine the responsive UI, diagnose live failures and create regression tests. The owner made product, privacy, service and release decisions. The current release has 313 passing offline tests across 50 files, 19 normal Chromium scenarios, strict TypeScript, zero-warning lint, an optimized production build and zero known production dependency vulnerabilities. The complete 20-scenario local-Supabase live browser suite passed against real Nemotron/Nebius and Tavily providers on October 5, 2026. The October 6 UI pass added saved-stage progress, a clarification handoff, clearer result and next-step guidance, collapsible sources and a Czech main flow.

Try it

Open the no-account fictional demo at https://avenli.vercel.app/demo and inspect a problem's Overview, Research, Options and Tasks. This demo uses predefined data and makes no live AI or search calls. For a saved provider-backed analysis, create an account at https://avenli.vercel.app. The public MIT-licensed source and setup instructions are at https://github.com/jkaczmarczyk96-code/resolve. Watch the 2:34 public demonstration at https://youtu.be/B1Aqo6qaYtc.