Proposal Engine

Production multi-agent pipeline that automates the full pre-sales cycle — discovery → scoping → pricing → proposal — across 6 business units and 22 services.

---

What it does

A pre-sales consultant spends 3–6 hours per proposal: discovery call notes, scoping, pricing lookup, formatting. This pipeline does it in under 90 seconds.

Input: client name, industry, size, region, service code. Output: a complete, client-ready proposal markdown — exec summary, scope, methodology, pricing table, assumptions, SLAs.

---

Architecture

Client Input (Streamlit UI or CLI)
         │
         ▼
 QuestionnaireAgent        ← generates discovery questions tailored to BU + client context
         │
         ▼
   DiscoveryAgent          ← extracts pain points, regulatory drivers, environment details
         │
         ▼
   ScopingAgent            ← estimates effort, defines in/out scope, flags risks
         │
         ▼
   PricingAgent            ← codebook-driven pricing with tier + discount logic
         │
         ▼
   CriticAgent             ← adversarial reviewer: challenges weak claims, vague scope, unsupported estimates
         │
         ▼
   ProposalAgent           ← renders full proposal from template, incorporating critic fixes
         │
         ▼
   Output: proposal.md     ← saved to outputs/{date}_{client}_{service}/

Model strategy: llama-3.1-8b-instant (MODEL_FAST) for agents 1–4 and Critic. llama-3.3-70b-versatile (MODEL_PRO) for ProposalAgent only. Keeps cost near zero on Groq free tier.

---

Services covered — 6 BUs, 22 codes

BU	Services
DFIR	IR Retainer (ENT/STD), Compromise Assessment, Internal Forensic Investigation, Breach & Attack Simulation, Digital Forensics
CTS (Pen Test)	Network PT, Web App PT, Mobile PT, API PT, Cloud PT, Red Team, Phishing Sim
GRC	ISO 27001, PCI DSS v4.0, SOC 2, HIPAA, DPDP Readiness
MXDR	Managed XDR Essential, Managed XDR Advanced
Institute (Training)	Security Awareness, DFIR Practitioner
DPG (Data Privacy)	DPIA, Data Mapping

---

Adversarial CriticAgent

Most multi-agent pipelines just chain outputs forward. This one includes a critic that runs against its own output:

# After exec summary is generated, critic challenges it before rendering
critique = critic.run(exec_summary, doc_type="executive summary")
# Returns: 3 weaknesses with exact quote + rewrite suggestion
exec_summary = critic.improve_exec_summary(exec_summary, critiques[:2])

Failure modes it catches: vague claims without evidence, unsupported timelines, missing risk acknowledgement, regulatory claims without citation.

---

Streamlit UI

Sidebar: API key status · pipeline mode (full/discovery/scope)
Main:    client info form → BU selector → service dropdown
         [Generate Proposal] → live phase progress
         Results tabs: Discovery · Scope · Pricing · Proposal · Downloads
         Pipeline history: last 10 runs, expandable, downloadable

Dark theme, real-time progress, download buttons for every output file.

---

Setup

Streamlit Community Cloud (recommended — free)

1. Fork this repo
2. Go to share.streamlit.io → New app → your fork → streamlit_app.py
3. Add secrets in the dashboard:

   GROQ_API_KEY = "gsk_..."
   COMPANY_NAME = "Your Company Name"

4. Deploy — auto-redeploys on every push to main

Local

git clone https://github.com/sanjayrkshetty/proposal-engine
cd proposal-engine
pip install -r requirements.txt

cp .env.example .env
# Edit .env: add GROQ_API_KEY=gsk_...

# Streamlit UI
streamlit run streamlit_app.py

# CLI
python main.py --service DFIR-R-ENT --client "Acme Bank" \
  --industry "Banking" --size "Mid-market" --region "India-South"

Environment variables

Variable	Required	Default
GROQ_API_KEY	Yes	—
COMPANY_NAME	No	Proposal Engine

---

CLI flags

--service    Service code (e.g. DFIR-R-ENT, CTS-NPT, GRC-ISO27001)
--client     Client name
--industry   Industry vertical
--size       Organisation size
--region     Region (India-South, SEA, MEA, North America...)
--mode       full (default) | discovery | scope
--tier       Standard | Enterprise (auto-detected if omitted)
--discount   Discount % (0–40)
--gam        GAM ID from config/gam_list.yaml

---

Tests

pytest tests/ -v
# 90/90 passing — zero LLM calls in CI (all mocked)

Tests cover: template rendering for all 22 service codes, placeholder completeness, proposal number format, billing contact propagation.

---

Output structure

outputs/
└── 2026-05-11_Acme-Bank_DFIR-R-ENT/
    ├── questionnaire.json
    ├── discovery_brief.md
    ├── scope_estimate.md
    ├── pricing.md
    └── proposal.md          ← the deliverable

---

Extending

Add a new service: edit config/services.yaml + add a template to templates/{bu}/.

Add a new BU: create bu/{name}/config.py and bu/{name}/questionnaire.py, register in bu/__init__.py.

Swap the LLM: change MODEL_FAST / MODEL_PRO in agents/base_agent.py. Any OpenAI-compatible API works (OpenRouter, Together, local Ollama).

---

Built by

Sanjay R K Shetty — AI Security Researcher, MIT Bengaluru '26.

© 2026 Sanjay R K Shetty · All rights reserved