Loan EMI Negotiator

An AI agent that turns an Indian borrower's loan profile into a lender-specific rate-negotiation artifact: target rate band, ranked arguments with sources, draft email, and phone script.

Live demo → loan-emi-negotiator.netlify.app

Most borrowers know they're probably overpaying. The hard part isn't knowing — it's writing the actual ask. This product collapses that last mile.

The problem

Indian retail borrowers typically pay 0.5–1.5% above their lender's own published rate floor for the same CIBIL band, simply because they never write the email asking for a reset. Generic finance chatbots explain repo rates and EBLR; the borrower still has to figure out what to say to the bank. The gap between informing and acting is where the real product value lives — and where trust is hardest.

The product wedge

Three deliberate scope choices that shape the whole product:

1. One painful workflow, not "AI personal finance". The agent does one thing: produce a credible, lender-aware rate-reduction request. Narrow scope is what makes the AI feel reliable.
2. Action artifact, not a chat. Output is a structured plan — target rate, ranked arguments, email, phone script — that the user can edit and send. No conversational ambiguity.
3. Trust design over model quality. Every claim cites a source. Assumptions and confidence are surfaced. Nothing auto-sends. The model is a drafter and strategist, not an authority.

What the user gets

A single artifact, generated in ~90 seconds:

• Target rate band mapped to the lender's published EBLR floor + the borrower's CIBIL band
• Estimated EMI reduction & total interest saved
• Ranked negotiation arguments, each with a public RBI / lender source URL
• Draft email to the relationship manager (editable)
• Phone script with three pre-handled objections
• Stated assumptions, confidence level, and missing inputs

Supports home, personal, auto, LAP, education, gold, and business loans across major Indian lenders (SBI, HDFC, ICICI, Axis, Kotak, BoB, PNB, etc.).

Key product decisions

Decision	Why
Hardcoded lender rate-card snapshot in lib/domain.ts	Grounds the model with verifiable floors. A live scrape would be more accurate but introduces audit/freshness risk for a portfolio MVP — explicitly flagged as a follow-up.
Optional statement upload (PDF/image) → Gemini extract	Removes the biggest drop-off — manual data entry — without making the doc upload a hard requirement. Users can still type values in.
Demo profiles + deterministic fallback	A live walkthrough cannot fail because of API quota. Reliability is a product feature when a recruiter or stakeholder is watching.
Streaming structured output (Zod schema)	Faster perceived latency and a typed contract that prevents malformed plans from rendering.
Editable email + phone script	The model drafts; the human commits. Treats the borrower as accountable, not as a passive consumer of AI output.

Reliability architecture (graceful degradation)

The agent is built to never show a broken state during a demo or a quota spike. The negotiate route walks down a fallback ladder:

DEMO_MODE=true ─────────► deterministic plan from lib/demo.ts            (no API call)
       │
       ▼
GEMINI_API_KEY missing ─► deterministic plan (if DEMO_FALLBACK on)       OR 500 error
       │
       ▼
gemini-2.5-flash-lite ─► generates plan ✓
       │ (rate-limit / 5xx)
       ▼
gemini-2.5-flash ──────► generates plan ✓
       │ (rate-limit / 5xx)
       ▼
gemini-2.0-flash-lite ─► generates plan ✓
       │ (rate-limit / 5xx)
       ▼
gemini-2.0-flash ──────► generates plan ✓
       │ (all four exhausted)
       ▼
DEMO_FALLBACK=true ────► deterministic plan (response tagged X-Demo-Reason)
       │ else
       ▼
429 rate_limited / 500 generation_failed (surfaced to UI)

Every fallback response carries an X-Demo-Reason header (demo_mode, missing_api_key, quota_fallback, provider_fallback) so it's auditable in the network tab.

Tech stack

Next.js 16 (App Router) · React 19 · TypeScript · AI SDK + Google Gemini · Zod schemas · React Hook Form · Tailwind CSS

Run locally

npm install
cp .env.example .env.local
# Add GEMINI_API_KEY (from Google AI Studio)
npm run dev

Open http://localhost:3000 → click Try the agent or go to /negotiate.

For a quota-proof local demo: set DEMO_MODE=true in .env.local and restart.

Deploy on Netlify

netlify.toml, .nvmrc, and .node-version are committed.

In the Netlify dashboard:

• Build command: npm run build · Publish dir: .next · Node: 20
• Set GEMINI_API_KEY (Functions + Builds scope, never NEXT_PUBLIC_)
• For a public portfolio link, also set DEMO_MODE=true to prevent quota drain by visitors
• Or set DEMO_FALLBACK=true for live AI with deterministic fallback on errors

What's not there yet

Intentional follow-ups, called out so the scope is honest:

• Lender rate-card data is a hand-curated snapshot, not a live scrape with audit trail.
• Generated advice is not financial advice — must be verified against the loan agreement.
• Statement extraction quality varies with PDF/image quality.
• No persistence, accounts, or encrypted document storage.
• No automated eval suite for hallucination, source quality, or plan usefulness.
• Non-home-loan pricing uses indicative benchmarks rather than fully scraped lender-specific cards.

Not financial advice

Snapshot data dated May 2026. Verify every rate, fee, and slab with your lender before acting.