Enterprise Data AI — Semantic Layer

Solon – Pre-sales AI agent

From a client RFP to a client-ready proposal, in minutes, not afternoons.

Proposal Intelligence reads a requirements document and returns a branded Word proposal, an Excel feature list, and an hours estimate grounded in projects we have already delivered. You review and approve each step before anything leaves your screen.

    By ticking this box, you agree to ⋮IWConnect’s Terms & Privacy Policy. You also agree to receive future communications from ⋮IWConnect. You can unsubscribe anytime.

    See the pipeline

    Proposal Intelligence pipeline

    Drop in a requirements document in PDF, DOCX, XLSX, MD, or TXT. In minutes, Proposal Intelligence returns a branded Word proposal, an Excel feature list with best, likely, and worst-case hours, an HTML evaluation report scored by two AI judges across five dimensions each, and a sha256-hashed append-only JSONL audit log. A human-in-the-loop checkpoint requires you to approve the hours estimate before any proposal copy is generated.

    12–24hrs
    Senior time saved per RFP
    What it takes today, reported by IWConnect consultants
    10
    Specialist agents in the pipeline
    One graph, one trace, one audit log
    2
    Independent AI judges per run
    5 dimensions each, runs that fail are blocked
    82%
    Test coverage on the code
    296 unit tests, ruff and mypy strict on every commit
    The problem

    Every new RFP blocks a senior for half a day.

    From first read-through to a sendable proposal, consultants spend 12 to 24 hours drafting, looking up past hours, finding services, and formatting. Junior team members cannot help, because the institutional knowledge of services, past projects and pricing lives in a few heads.

    Today, one opportunity, end-to-end
    A senior spends the morning on one proposal.
    • 0:00 Read the brief
    • 0:30 Map to services
    • 1:30 Hunt for past hours
    • 3:00 Estimate and review
    • 4:30 Draft and format the documents
    Net cost per opportunity
    ½ day of senior time, gone.
    Quality varies by person. Response time varies by workload.
    Senior consultant time
    12 to 24 hours, every time.

    Drafting, looking up past hours, finding services, formatting. That is what consultants told us when we asked. One person, every time.

    The real bottleneck
    Pre-sales competes with delivery for the same people.

    The pre-sales pipeline pulls senior consultants away from the projects they are running. When the queue grows, response times slow, and consistency drops with it.

    The solution

    In, out, and a human in the loop.

    Drop a requirements document and get four artifacts back. The branded proposal you can send, the estimate behind it, an evaluation report from two independent judges, and an audit log of every decision the pipeline made.

    .docx

    Branded Word proposal

    Branded template, fully populated and ready to send.

    .xlsx

    Excel feature list

    Feature-by-feature hours breakdown: best, likely, worst.

    .html

    Evaluation report

    Two independent AI judges, five dimensions each.

    .jsonl

    Audit log

    Every decision, sha256-hashed and append-only.

    HITL checkpoint

    You approve the estimate before a single word is written.

    The graph pauses after estimation. Accept the numbers, or redirect to any upstream agent and re-run from there.

    After estimation, before generation
    Architecture · 10 nodes, one graph, one trace

    Each agent owns one decision. Two independent judges score the result.

    Proposal Intelligence is not one model with a complicated prompt. It is a pipeline of small, focused agents, each with a single job. They run in parallel where they can, in sequence where they must, and every step is observable in real time.

    00 · pre-step
    Ingestion

    Azure Doc Intelligence, with pdfplumber and pypdf as fallbacks. python-docx for Word. XLSX, MD and TXT parsed natively.

    01 · guard
    Guard

    Scope check. Reject anything outside the company's portfolio so the pipeline does not hallucinate or burn tokens on it.

    02 · summary
    Summary

    Extracts context, requirements, constraints and red flags with severity ratings.

    03 · solution
    Solution

    Maps each requirement to a real company's service. No invented offerings.

    04 · estimation
    Estimation

    Hours per feature, calibrated against quality-filtered past runs with a score of at least 0.7.

    ⚡ hitl
    You approve

    Graph pauses. Review the estimate. Accept it, or redirect to any upstream agent to re-run.

    05 · features
    Features

    Structured deliverable list. What we would build, in what order, with dependencies.

    06a · parallel
    Word

    Branded .docx in the company's template.

    06b · parallel
    Excel

    Branded .xlsx with hours and complexity per feature.

    07a · judge
    Est. Judge

    Hours realism, overhead completeness, confidence calibration.

    07b · judge
    Prop. Judge

    5-dimension proposal score. Can hard-fail the run.

    every node
    A2A trace

    agent_id, model, tokens in and out, confidence, trace_id, sha256 output hash. Written to JSONL on every step.

    A2A message →
    agent_id: "estimation" model: "openai/gpt-5.1" tokens: 6,651 in, 2,327 out confidence: 0.80 output_hash: sha256 · 5bc96d7c…
    Built with AI · Spec-first, skills enforced, measured

    Right model per agent. Rules enforced mechanically, not by convention.

    Three models, one routing layer, and a development workflow where every agent was scaffolded the same way. Conventions live in skills and a shared contract file, so seven team members write code that looks like the same person wrote it.

    Built with

    Claude Code, Opus and Sonnet

    AI pair programmer for every agent, every session, every team member. Same skills, same review pattern, same conventions across the team.

    Pair programmer for all 9 agents
    Lighter agents

    GPT-5.4-mini

    Summary, solution and mediator agents. Fast and cost-efficient where the reasoning surface is small.

    Routed via LiteLLM, swap per agent
    Heavy reasoning

    GPT-5.1

    Estimation, features, proposal and evaluation. The agents where calibration and judgement matter most.

    One config line to change any agent
    Want a Demo?

    Bring an RFP. We will run it live.

    Drop in one of your real requirements documents. Watch ten agents run end-to-end, approve the estimate, and open the branded .docx and .xlsx your team would send. Make no mistakes, with a paper trail every enterprise client can audit.

      By ticking this box, you agree to ⋮IWConnect’s Terms & Privacy Policy. You also agree to receive future communications from ⋮IWConnect. You can unsubscribe anytime.