1,247 pages. PDF. Scanned. He missed the eligibility criteria on page 983. Bid rejected.

That night I started building Tender AI.

The Problem Nobody Talks About

Indian MSMEs lose ₹2.3 lakh crores worth of government contracts every year not because they're unqualified — but because they:

1. Find tenders too late (checking GeM, CPPP, NHAI manually)

2. Can't read fast enough (average tender: 400-1,200 pages)

3. Miss hidden clauses (payment terms on page 742, LD on page 891)

The industry win rate? 15%.

What I Built: An 8-Agent Autonomous Workforce

Instead of one chatbot, I built 8 specialized AI agents that work like a real bid team:

1. Discovery Agent (Playwright crawler)

• Checks GeM, CPPP, NHAI, Smart Cities every 60 minutes

• Found 23 relevant tenders last month automatically

2. Eligibility Agent

• Compares tender requirements vs your company profile

• Example: "ISO 27001 required? You have it. Turnover >₹2Cr? You have ₹3.5Cr."

3. Research Agent (Azure AI Search + RAG)

• This is the magic. It doesn't "summarize" — it grounds every answer

• Ask: "What's the payment term?" → Returns exact page 742, clause 12.3

4-8. Competitor Intel, Planner, Proposal Generator, Execution, Reflection

• Generates the actual technical bid

• Stages it in your Outlook

• Validates citations before sending

Real Numbers From Last 30 Days

I tested this with my friend's company:

• Autonomy Score: 91% (up 2.5%)

• Tenders Evaluated: 23 (0 manual hours)

• Proposals Generated: 7

• Target Capture Rate: 43.4% (vs 15% industry)

One tender: "Implementation of Real-time AI Operations & Big Data Analytics Dashboard"

• CPPP published it June 8

• My AI found it at 2:14 AM

• Match Score: 92%

• Win Probability: 76%

• Proposal ready by 2:26 AM

He was sleeping.

The Tech Stack (for builders)

• Orchestration: Semantic Kernel (Microsoft)

• Crawling: Playwright

• RAG: Azure AI Search + GPT-4o

• Database: Cosmos DB (NoSQL + Graph for competitor relationships)

• Integration: Microsoft Graph API (Outlook + Planner)

• Hosting: Render (free tier for now)

Total build time: 6 weeks. Cost so far: ₹0 (using free tiers).

Why This Works for India Specifically

Generic tools like ChatGPT fail on Indian tenders because:

• They hallucinate on scanned PDFs

• Don't understand "EMD", "LD clause", "GeM bid", "two-bid system"

• Can't parse Hindi + English mixed documents

I trained the model on 200+ successful Indian electrical and IT sector bids. It knows what a "Form 26AS" is.

Current Status: Beta

I'm giving free access to 50 Indian MSMEs who are actively bidding on GeM/CPPP.

What you get:

• Automated monitoring of 4 portals

• Instant match scoring

• Proposal drafts (you approve before sending)

What I need from you:

• Feedback on accuracy

• Tell me which portal I missed

Try the live demo: tenderai.publicvm.com

Or DM me on LinkedIn I'll set up your profile manually.

What's Next?

Next post: I'll break down exactly how the RAG pipeline extracts clauses from 1,200-page PDFs without hallucinating (and why ChatGPT-5 fails at this).

Follow for build-in-public updates. Building this solo, documenting everything.

Building TenderPilot AI | Autonomous bid workforce for Indian MSMEs Connect:
saultytkc13@gmail.com