tendr.bid
All docs
Reference

AI (QVAC)

How the three AI buttons in the app work, where data flows when you use them, and what the privacy story is — without any third-party AI provider in the loop.

View on GitHub

AI on tendr.bid (powered by QVAC)

Tendr ships three AI surfaces — RFP scope drafting, bid drafting, and post-decrypt bid comparison — using Tether QVAC as the inference engine. Same workflow shape as ChatGPT-style "draft this for me", but the bytes never leave for OpenAI / Anthropic / Google. They go from your browser straight to a QVAC sidecar we operate ourselves.

Where the AI shows up

SurfaceWho uses itWhat it does
RFP create → Draft with AI button next to Scope summaryBuyerType a paragraph describing what you need; QVAC returns a structured scope (objectives, deliverables, milestones, success criteria) you can edit before posting.
Bid composer → Start drafting bid with AI button next to ScopeProviderReads the RFP scope + your optional context (tech stack, target price, timeline). Returns a complete bid draft — price, timeline, scope markdown, and a populated milestones array with acceptance criteria. Every field of the bid form gets filled in on accept.
Award screen → Compare bids with AI button under the decrypted bid listBuyerAfter you've decrypted the sealed bids, QVAC generates a side-by-side comparison table + recommends a winner with reasoning.

All three buttons hide themselves automatically when the QVAC sidecar URL isn't configured (NEXT_PUBLIC_QVAC_BASE_URL). The rest of the product works identically without AI — nothing in the bidding, awarding, escrow, milestone, or reputation flows depends on AI being present.

How it works

  1. Your browser calls the QVAC sidecar directly over HTTPS.
  2. The QVAC sidecar is a small process running QVAC on the Bare runtime, exposing an OpenAI-compatible HTTP server (POST /v1/chat/completions). It loads the Qwen3 4B Instruct (Q4_K_M quant) model into a single dedicated GPU and runs inference there.
  3. The response comes back to your browser. The sidecar holds no per-user state — request in, response out.
Your browser  ───────HTTPS──────▶  QVAC sidecar (Nosana GPU)
              ◀──────HTTPS──────
              (decrypted bids)        (Qwen3 4B Instruct Q4_K_M
              (RFP scope, etc.)        via QVAC's OpenAI-compatible server)

What's not in this path:

  • Tendr's Next.js server (we never see the prompts or responses).
  • Any closed AI provider (OpenAI, Anthropic, Google, etc.).
  • Any multi-tenant inference API.

Why "OpenAI-compatible" doesn't mean OpenAI

QVAC's serve mode speaks the same wire format as OpenAI's official API (POST /v1/chat/completions with the same JSON schema). That's so we can use the standard openai npm package on the client without writing custom QVAC code. The bytes don't leave for openai.com. The protocol shape is mimicked; the destination is our sidecar's URL.

Same idea as "S3-compatible storage" — the API is mimicked but the bytes go to whoever's storage is configured, not to AWS.

The privacy story (what we actually deliver)

Here's the architecture, with no embellishment:

  • No closed AI providers in the pipeline. The model is open-weight (Qwen3 4B Instruct, Q4_K_M) and the inference engine is QVAC. No OpenAI API key, no Anthropic API key, no third-party AI vendor's data-collection policy applies.
  • Tendr backend never sees AI data. The browser hits the Nosana endpoint directly via the env var NEXT_PUBLIC_QVAC_BASE_URL. There is no Tendr API route that proxies the call. Our Vercel servers — where wallet sessions, RFP metadata, and bid index live — never see prompts, decrypted bid plaintexts, or the model's response. Verifiable by reading apps/web/lib/ai/client.ts — the OpenAI client's baseURL points at the Nosana endpoint, not at any /api/* Tendr path.
  • Single-tenant inference. The Nosana GPU is a dedicated container running our QVAC image. Not a multi-tenant inference API serving 50 other apps. No prompt-mixing.

What we are NOT claiming:

  • ❌ NOT local-first inference for end users. The model runs on a Nosana GPU we operate, not on your laptop. Running a 4B-parameter model on a buyer's browser isn't realistic today.
  • ❌ NOT zero trust. You're trusting our Nosana deployment instead of trusting OpenAI's data policies. Different threat model, real privacy improvement, not "no trust required".
  • ❌ NOT an escape hatch for individual users to swap to their own local QVAC. Tendr is a hosted consumer app — env vars are baked into the Vercel build, end users can't reroute the AI calls without forking and self-deploying the entire app.

Comparison vs. the alternatives

SetupWhere bid plaintext livesPrivacy strength
Browser → OpenAI APIBuyer's machine + OpenAI's data centers + their training set if you don't opt outWeakest
Browser → Tendr backend → some AI providerBuyer's machine + Tendr's servers + the AI provider's serversWorse than today's setup
Browser → QVAC sidecar on Nosana (today's setup)Buyer's machine + our dedicated Nosana nodeStrong — open-weight model on infra we run; no closed AI provider; Tendr backend out of the loop

What the AI is good at + not good at

Good at: structured tasks with clear inputs and a known output shape.

  • Drafting RFP scopes from a few sentences of intent
  • Drafting starting-point bids that respect the provider's stated context (tech stack, target price, timeline)
  • Comparing decrypted bids on observable dimensions (price, timeline, scope coverage, milestone realism, risk flags)

Not good at:

  • Replacing your judgment about counterparties — the recommended winner is a starting point, not a decision
  • Niche technical evaluation outside the model's training data
  • Pricing on rare specialties (the model's training data on, say, M&A advisory or quant trading audits is uneven)

The right mental model: a structured-output assistant. Treat its output as a draft you'll always edit.

What we don't do

  • No AI signing. The model never signs a transaction, never moves money, never touches the on-chain side. It only reads inputs and returns text.
  • No required AI for product behavior. Every flow that has an AI button works identically without it — bidding, awarding, escrow, milestones, reputation are all independent.
  • No conversation history. Each AI call is a one-shot completion. The sidecar holds no per-user state between requests.

Reference

More docs