PadippuAI — Building Tamil Nadu's First Open Engineering-Admissions MCP Server with Amazon Nova Pro
Most Tamil Nadu students miss 80% of scholarships they qualify for and choose engineering colleges blind to placement data. This post walks through building a bilingual English + Tamil AI agent on Amazon Nova Pro that unifies TNEA cutoffs, scholarships, NIRF ranks, fees and career pathways behind a custom MCP server — reusable by Claude Desktop and any future agent.
Engineering admissions in Tamil Nadu generate enormous amounts of information every single year — TNEA cutoffs, NIRF rankings, college fee structures, hostel costs, scholarship deadlines, placement reports, career pathways, and government schemes.
However, most students and parents still access these insights through fragmented sources.
Cutoffs come from one portal. Scholarships from another. NIRF ranks from a third. Fees and placement data are scattered across college websites. Career guidance often depends on word-of-mouth.
I wanted to explore whether an AI Agent could unify these workflows into a single intelligent admissions assistant.
This led to the development of PadippuAI ("Padippu" = education in Tamil) — a smart college-counselling assistant built using Amazon Nova Pro, the Strands Agents SDK, a custom-built Model Context Protocol (MCP) server, persistent student memory, structured TN-education datasets, and bilingual English + Tamil output for parents.
The project was developed as part of the AWS User Group Madurai – Nova Month Builders Skill Sprint 2026, Challenge 5 (Innovate).
Problem Statement
Students writing TNEA every year repeatedly face the same challenges:
- Difficulty identifying which colleges are realistically within reach
- Limited visibility into NIRF rankings and placement data for tier-2/3 colleges
- Lack of awareness about scholarships they are eligible for
- Confusion about 4-year cost (tuition + hostel) and how much scholarships actually reduce it
- No clarity on which colleges are near home in their own district
- Fragmented career-guidance information after the B.E. — higher studies, government exams, industry roles
The objective of PadippuAI was not simply to build another chatbot.
The objective was to create an admissions intelligence system capable of combining:
Structured data + agent reasoning + memory + decision support + reusable tool surface (MCP).
And critically — to make the tool surface reusable by any other agent, not locked inside one app.
System Architecture
PadippuAI follows an agent-based architecture built on the Strands Agents SDK with a custom FastMCP server as the tool layer.
┌─────────────────────────────────────────────────────────────┐
│ PadippuAI Agent │
│ │
│ Strands Agent (model = us.amazon.nova-pro-v1:0) │
│ │ │
│ ├── system prompt (bilingual + workflow rules) │
│ ├── persistent student memory (JSON file) │
│ └── tools loaded over MCP stdio │
│ │ │
└──────────────────────────────┼──────────────────────────────┘
│ stdio (MCP)
▼
┌─────────────────────────────────────────────────────────────┐
│ mcp_server.py (FastMCP) │
│ │
│ cutoff │ search_colleges list_branches compare │
│ schemes │ list_scholarships get_scholarship │
│ location │ nearby_colleges │
│ NIRF │ nirf_top_colleges get_nirf_rank │
│ │ get_placement_stats │
│ ROI │ estimate_total_cost │
│ careers │ career_pathways │
│ meta │ dataset_info │
└─────────────────────────────────────────────────────────────┘
│ reads
▼
data/tnea_cutoffs.csv, data/scholarships.json,
data/nirf_engineering_2024.csv, data/college_fees.csv,
data/college_locations.csv, data/placement_stats.csv,
data/career_pathways.jsonThis architecture lets PadippuAI combine structured admissions intelligence with persistent student context — and exposes that intelligence as a reusable MCP tool surface that any other agent (Claude Desktop, future Strands apps) can plug into.
The Conversation Layer
The very first iteration was a plain Nova Pro chatbot with no tools. It could talk about colleges but had no real data. Students would ask "Can I get TCE Madurai with 178?" and the model would hallucinate a confident but wrong number.
That was the moment that made it obvious: tools matter more than the model.
Cutoff Intelligence
The first real capability was a TNEA cutoff search.
Example:
I scored 480/600 (BC). Which CSE colleges am I eligible for?PadippuAI converts marks (TNEA = (480/600) × 200 = 160.0), then calls:
search_colleges(min_marks=160, category="BC", branch_code="CSE", limit=10)Output:
ENGLISH SUMMARY
• 12 CSE colleges match. Top by cutoff: SSN (197.0), PSG (196.5), TCE (195.0).
• Within Madurai: TCE (195.0), Anna Univ Regional Madurai (193.0).
• Safe option: Latha Mathavan Madurai (151.0).
தமிழ் சுருக்கம்
உங்களுக்கு 12 CSE கல்லூரிகள் கிடைக்கும். மதுரையில் TCE மற்றும்
Anna Univ Regional இரண்டுமே சிறந்த தேர்வுகள்.
ACTION
Visit TCE Madurai open day before TNEA choice filling.The bilingual format wasn't decorative — parents in tier-2 towns read the Tamil block first.
Two more cutoff tools were added in the same layer:
-
list_branches_for_college→ "What does college code 4 offer?" returns the full cutoff matrix across all six branches and six categories.
-
compare_colleges→ "Compare 1, 7, 8 for CSE BC" returns a side-by-side row for each.
Persistent Student Memory
After every conversation reset, the student had to repeat their marks, category, district, and income. That broke immersion.
Inspired by AgriNova's use of Mem0 + FAISS, I added a tiny file-based memory layer inside
starter.py that persists the same shape (profile + shortlist + applied scholarships) but uses a single memory.json file instead of a vector store. The workshop runs offline with zero extra installs.Capture is deterministic and heuristic-based. When the student types:
I scored 480/600, BC, from Madurai.The starter prints:
[memory] remembered twelfth_total=480/600 -> tnea_marks=160.0
[memory] remembered category=BC
[memory] remembered home_district=MaduraiAnd from that point on, the agent's system prompt carries a compact
[REMEMBERED STUDENT PROFILE] block. Future questions like "what scholarships am I eligible for?" no longer need to repeat the category — the agent fills it from memory before calling list_scholarships ./memory shows what's stored. /reset wipes it.
Government Scholarship Discovery Engine
Most TN students are unaware that they are eligible for stackable scholarships — central + state + AICTE + minorities welfare schemes can sometimes cover 100% of tuition together.
PadippuAI ships a curated
scholarships.json containing schemes from:Central
- PMS-SC / PMS-ST / PMS-BC (NSP — Ministry of Social Justice)
- AICTE-PRAGATI (girl child)
- AICTE-SAKSHAM (specially-abled)
- Central Sector Scheme for College & University Students
- Inspire Scholarship (Science stream)
Tamil Nadu
- TN Free Education for First-Generation Graduates
- TN Minorities Welfare Scholarship
- Adi Dravidar Welfare Special Scholarship
- TN Most Backward Classes Welfare scholarship
Example:
What scholarships am I eligible for?Since the student already told PadippuAI they are BC, female, income 1.8 lakh, the agent silently calls:
list_scholarships(category="BC", income_below=180000, gender="F")Output:
ENGLISH SUMMARY
• PMS-BC: ₹48,000/year, deadline 31 Aug.
• AICTE-PRAGATI: ₹50,000/year, girls-only, deadline 31 Oct.
• TN-FREE-EDU-FIRST-GRAD: full tuition waiver (~₹40-90k), apply
via college on admission.
தமிழ் சுருக்கம்
உங்களுக்கு 3 உதவித்தொகை கிடைக்கும் — இரண்டும் சேர்த்து வருடம்
₹98,000 வரை. முதலில் முதல் தலைமுறை பட்டதாரி திட்டத்தை
கல்லூரி அலுவலகத்திலேயே விண்ணப்பிக்கவும்.
ACTION
Open scholarships.gov.in this week. Aadhaar-linked bank account ready.get_scholarship("AICTE-PRAGATI") returns the full record — eligibility rules, document list, exact URL.
Location / District Intelligence
A 480/600 student doesn't necessarily want to leave home. PadippuAI added a district-aware tool:
What CSE colleges are near Madurai?nearby_colleges(district="Madurai", branch_code="CSE", radius_km=100)This first looks for an exact district match; if none, it falls back to a haversine radius search using OpenStreetMap-derived district centroids. Each result includes the nearest railway station and bus route — practical decision data parents care about.
NIRF + Placement Intelligence
Cutoffs answer "can I get in". They don't answer "should I want to".
Two more tools answer the second question:
NIRF Rankings
nirf_top_colleges(state="Tamil Nadu", limit=10)
get_nirf_rank("Thiagarajar College")This pulls from the official NIRF Engineering 2024 list (IIT Madras at #1 national, TCE at TN rank-X, etc.) bundled as
nirf_engineering_2024.csv .Placement Statistics
get_placement_stats(college_code="7", branch_code="CSE")Output:
{
"college_name": "Thiagarajar College of Engineering",
"branch_code": "CSE",
"avg_package_lpa": 9.0,
"median_package_lpa": 7.0,
"highest_package_lpa": 28.0,
"placement_percent": 92,
"top_recruiters": ["Microsoft", "Amazon", "Zoho", "Freshworks", "Visa"]
}The median is deliberately reported alongside the average — students were being misled by single "highest package" numbers that came from one outlier offer.
Education ROI (Farm-Economics Analogue)
The AgriNova project had a Farm Profit Calculator. PadippuAI's analogue is the 4-year B.E. cost estimator:
estimate_total_cost(college_code="7", hosteller=true, scholarship_inr_per_year=48000)Output:
ENGLISH SUMMARY
• TCE Madurai — Aided college.
• Annual: tuition ₹85,000 + hostel ₹55,000 = ₹1,40,000 gross.
• PMS-BC reduces it to ₹92,000/year.
• Four-year net cost ≈ ₹3,88,000 (vs ₹5,80,000 without scholarship).
• Savings from PMS-BC over 4 years: ₹1,92,000.
தமிழ் சுருக்கம்
PMS-BC எடுத்தால் நான்கு ஆண்டுக்கு ₹1.92 லட்சம் சேமிக்கலாம்.
ACTION
Get caste + income certificates this month. Apply via scholarships.gov.in
the same week TNEA result is announced.This turned the assistant from "where do I get in" into "what is the real cost-after-aid" — the question that actually decides college selection in middle-income families.
Knowledge Layer (CSV + JSON)
One major design call: no hardcoded logic in the MCP server. All admissions data lives in seven flat files in
data/ :File | Rows | Source |
|---|---|---|
tnea_cutoffs.csv | 60 | TNEA official portal (2024 Round-1) |
scholarships.json | 10 schemes | scholarships.gov.in + AICTE + TN Welfare |
nirf_engineering_2024.csv | 25 | nirfindia.org |
college_fees.csv | 30 | college sites + TN Fee Committee |
college_locations.csv | 30 | OpenStreetMap district centroids |
placement_stats.csv | 62 | college placement reports (2023-24) |
career_pathways.json | 6 branches | AICTE Career Compass + TNPSC notifications |
A volunteer can update next year's cutoffs by editing a CSV — no Python changes required. This was the AgriNova lesson restated: hardcoded guidance → structured datasets.
Career Pathways
A student getting CSE doesn't just want to know about colleges. They want to know what happens after the B.E.
What careers are possible after CSE?career_pathways(branch_code="CSE")Returns three structured branches of advice:
- Core industry roles — Software Engineer, Data Engineer, ML Engineer, Cloud Engineer, Cybersecurity Analyst, DevOps/SRE.
- Higher studies — M.Tech via GATE / MS abroad (GRE) / MBA via CAT.
- Government exams — ISRO Scientist 'SC', DRDO RAC, BARC OCES, UPSC IES (IT), Banking IT Officer (IBPS SO).
- Entrepreneurship hint — SaaS / fintech / agritech (the local Zoho-style ARR play).
The same tool exists for ECE, EEE, MECH, CIVIL, IT — each curated separately because the government-exam route for a MECH student (TNPSC AE Mechanical, BHEL/NTPC GET, Railway JE) is completely different from the route for a CSE student.
The Custom MCP Server (the Innovation)
This is the deliverable that makes PadippuAI a Challenge 5: Innovate project.
Why MCP and not just Python tools?
A normal Strands agent would import each tool as a Python function inside the same process. That works — but it locks the tool surface inside one app.
By packaging the same tools as an MCP server, the entire TN-education tool surface becomes:
- Reusable by Claude Desktop (drop into
claude_desktop_config.json)
- Reusable by any future Strands agent (different prompt, same tools)
- Reusable by VSCode Copilot, Cursor, future MCP clients
- Reusable by other AWS User Group projects that need TN-education data
This is what the workshop calls "infrastructure for the next agent" — every team building an admissions assistant doesn't have to rebuild scrapers and rule encoders. They
pip install and connect.How the agent connects
server_params = StdioServerParameters(
command=sys.executable,
args=[str(SERVER_SCRIPT)],
env={**os.environ},
)
with MCPClient(lambda: stdio_client(server_params)) as mcp_client:
tools = mcp_client.list_tools_sync() # all 12 tools discovered
agent = Agent(model=MODEL, tools=tools, system_prompt=build_system_prompt())MCPClient spawns the server as a subprocess, lists the tool catalog over stdio, and hands the typed schema to Strands — which exposes them to Nova Pro as native tool calls.How to reuse it from Claude Desktop
{
"mcpServers": {
"tn-education": {
"command": "python",
"args": ["/abs/path/to/mcp_server.py"]
}
}
}That's the whole installation. Claude Desktop now has the same 13 TN-education tools available globally.
Technology Stack
Layer | Technology |
|---|---|
Agent Framework | Strands Agents SDK |
LLM | Amazon Nova Pro ( us.amazon.nova-pro-v1:0 ) |
Tool Protocol | Model Context Protocol (FastMCP) |
Persistence | File-based JSON (Mem0/FAISS analogue) |
Data Layer | CSV + JSON (TNEA, NIRF, fees, placements, locations, careers, schemes) |
Language | Python 3.10+ |
Locale | English + Tamil bilingual |
PadippuAI Capability Layers
Capability Layer | Key Addition |
|---|---|
Conversation | Nova Pro baseline |
Cutoff Intelligence | TNEA 2024 search + branch matrix |
Persistent Memory | Student profile carried across sessions (Mem0/FAISS analogue) |
Scholarship Discovery | NSP + AICTE + TN welfare schemes |
Location Intelligence | Nearby colleges + haversine radius |
NIRF + Placements | National ranking + recruiter stats |
Education ROI | 4-year net cost after scholarship |
Knowledge Layer | All admissions data externalised to CSV / JSON |
Career Pathways | Roles + higher studies + govt exams |
Custom MCP Server | Reusable TN-education tool surface — the innovation |
Example Session (end-to-end)
============================================================
PadippuAI — Challenge 5: Custom TN-education MCP server
MCP server: mcp_server.py (12 tools)
Memory file: memory.json
============================================================
[MCP connected — 12 tools loaded from tn-education]
You: I'm Priya, scored 480/600, BC, girl, from Madurai. Family income 1.8 lakh.
[memory] remembered twelfth_total=480/600 -> tnea_marks=160.0
[memory] remembered category=BC
[memory] remembered gender=F
[memory] remembered home_district=Madurai
[memory] remembered family_income_inr=180000
PadippuAI: Hi Priya! Saved your profile. Want me to start with eligible
CSE colleges in Madurai?
---
You: Yes — show me CSE colleges I can get.
[tool: search_colleges(min_marks=160, category=BC, branch_code=CSE)]
ENGLISH SUMMARY
• 12 CSE colleges match for BC at TNEA 160.
• In Madurai: TCE (cutoff 195) — out of reach.
• Reachable in Madurai: Latha Mathavan (151), Solamalai (142).
• Best stretch options outside Madurai: KLN Sivaganga (184) — too high.
தமிழ் சுருக்கம்
மதுரையில் இரு கல்லூரிகள் கிடைக்கும். வெளியூரில் சிவகங்கை
KLN-க்கு உங்கள் மதிப்பெண் போதாது.
ACTION
Visit Latha Mathavan campus open day this Saturday.
---
You: What scholarships am I eligible for?
[tool: list_scholarships(category=BC, income_below=180000, gender=F)]
ENGLISH SUMMARY
• PMS-BC: ₹48,000/year, deadline 31 Aug.
• AICTE-PRAGATI: ₹50,000/year for girl students, deadline 31 Oct.
• TN-FREE-EDU-FIRST-GRAD: full tuition waiver via college.
ACTION
Aadhaar-linked SBI/Indian Bank account before 15 July.
---
You: Estimate 4-year cost at Latha Mathavan with PMS-BC scholarship.
[tool: estimate_total_cost(college_code=28, hosteller=true, scholarship_inr_per_year=48000)]
ENGLISH SUMMARY
• Latha Mathavan — Self-Financed.
• Annual: tuition ₹55,000 + hostel ₹28,000 = ₹83,000 gross.
• Net after PMS-BC: ₹35,000/year.
• Four-year net: ₹1,50,000.
தமிழ் சுருக்கம்
PMS-BC சேர்த்தால் நான்கு ஆண்டுக்கு ₹1.5 லட்சம் மட்டுமே செலவாகும்.
ACTION
Apply on scholarships.gov.in the same day TNEA admission letter is received.
---
You: What careers after CSE?
[tool: career_pathways(branch_code=CSE)]
ENGLISH SUMMARY
• Industry: Software Engineer, Data Engineer, ML Engineer, Cloud, DevOps, Cybersec.
• Higher studies: M.Tech via GATE / MS in US / MBA via CAT after 2-3 yrs.
• Govt exams: ISRO Scientist 'SC', DRDO RAC, UPSC IES (IT), BARC OCES, IBPS SO.
• Startup hint: SaaS / fintech (Zoho-style local play).
ACTION
Start LeetCode + one cloud cert (AWS Cloud Practitioner) in 1st year.
---
You: Where does your data come from?
[tool: dataset_info()]
PadippuAI:
Cutoffs: TNEA official portal, 2024 Round-1. 30 colleges, 6 branches.
Scholarships: scholarships.gov.in + AICTE + TN Minorities Welfare. 10 schemes.
NIRF: nirfindia.org 2024 engineering category.
Placements: 2023-24 batch college reports.
Disclaimer: static dataset — verify deadlines on official portals.
Lessons Learned
PadippuAI evolved significantly during development. The major transitions:
- Hardcoded answers → structured datasets — externalising the knowledge layer was the biggest inflection point.
- Single-app tools → reusable MCP surface — packaging the tool surface as MCP was the actual innovation.
- Stateless chatbot → persistent student memory — adding the memory layer changed the UX completely.
- English-only → bilingual English + Tamil — parents in tier-2 towns read the Tamil block first.
- "Can I get in" → "should I want to + what is the cost-after-aid" — adding NIRF, placements and ROI closed the decision loop.
The single biggest engineering lesson: shipping tools as MCP, not as
import statements , costs almost nothing extra and unlocks every future agent. That is the real win of Challenge 5.
Future Roadmap
- TNEA round-2/round-3 cutoff history for trend analysis
- Hostel + mess fee transparency portal
- College alumni LinkedIn integration for real placement reach
- Voice interaction in Tamil for parents who can't type
- Publish
tn-education-mcpto the official MCP server registry
- Convert to a hosted HTTP MCP so phone apps can connect
- Extend to Polytechnic and Arts & Science admissions
🎓 Final Thoughts
PadippuAI started as a simple TNEA cutoff lookup and gradually evolved into an admissions intelligence platform integrating AI agents, persistent memory, a custom MCP server, structured TN-education datasets, and bilingual decision support.
The biggest takeaway from this journey was to start simple, iterate continuously, and externalise complexity into datasets and reusable tool surfaces.
This project provided practical experience in:
- Building production-grade AI agents with Strands + Nova Pro
- Authoring custom MCP servers (FastMCP)
- Designing persistent memory for stateful conversations
- Bilingual UX for first-generation students and their parents
- Translating fragmented government data into a reusable agent tool surface
Built as part of the AWS User Group Madurai – Nova Month Builders Skill Sprint 2026 🎓🇮🇳
Thank you for taking the time to read! If you found it helpful, please like, share, and tell a TN student about it. I'd love your feedback in the comments.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article