🌟 Inspiration

As a team, we’ve worked with many SaaS tools — project management, design, CRMs, cloud services.
And every few months, the same dreaded emails arrived:

“Your subscription is renewing at \$500/month.”

We knew negotiating could save money, but most of the time we either ignored it (too busy) or accepted the price (too tired to haggle).
It was frustrating to realize we were leaving money on the table simply because negotiation is a chore.

That’s when the idea hit us: what if an AI agent could handle this for us?
Something that doesn’t get tired, doesn’t get awkward, and always has a negotiation strategy.
That’s how Haggle.ai was born.

🤖 What it Does

Haggle.ai is like your AI procurement teammate that:

Reads incoming vendor renewal or quote emails
Thinks through negotiation strategies (polite ask, firm counter, creative term swaps)
Drafts multiple counter-offer emails for you to pick from
Handles the back-and-forth with your approval in the loop
Updates a savings dashboard so you can see how much you’ve saved

💡 Tagline: The AI that literally puts money back in your pocket.

Example:
A vendor renewal at \$500 → Haggle.ai negotiates → final price = \$400 →
You saved: $$ \$500 - \$400 = \$100/month = \$1200/year $$ 🎉

🛠️ How We Built It

We combined open models + smart orchestration:

Reasoning & Drafting: gpt-oss-20B (via Hugging Face Transformers)
Fine-tuning: Hugging Face PEFT (LoRA) on negotiation dialogues
Acceleration: NVIDIA TensorRT for fast inference + NeMo Guardrails for safe, professional tone
Local-first Serving: Ollama for offline runs and vLLM for efficient serving
Interface: FastAPI backend + lightweight React/Streamlit dashboard
Memory: SQLite + FAISS to track past deals, pricing benchmarks, and context

🚧 Challenges We Faced

Getting the tone right (early drafts sounded robotic or too aggressive)
Teaching the model to use real negotiation tactics rather than generic text
Running a 20B model locally without long delays
Building a UX where users still feel in control (human-in-the-loop approval was key)

🏆 Accomplishments

Built a multi-strategy debate: “Polite Agent” vs “Firm Agent” draft counters, orchestrator picks the best
Demoed a full flow: Vendor started at \$500 → Haggle.ai countered → Vendor dropped to \$400 →
Savings: $$ \$1200/year $$
Integrated OpenAI, Hugging Face, NVIDIA, Ollama, vLLM into one workflow
Built a savings ledger dashboard that makes ROI visible

📚 What We Learned

Open-weight LLMs can be customized to specific use cases with light fine-tuning
Strategy matters — giving agents roles (e.g., firm vs polite) produces stronger results
NVIDIA TensorRT + vLLM made large models feasible in real time
Simplicity in UX drives trust in autonomous agents