"Building FinOptix-14B: A FinOps Architect in 14 Billion Parameters"

Carlos Cortez

ccortezb

Follow

"How I fine-tuned a small language model to audit Terraform, parse AWS costs, and enforce cloud governance — in 40 minutes on a single GPU."

tags: finops, aws, huggingface, machinelearning

![ finoptix-14b-finops-slm](

The Problem

I run a small AWS-based startup. Every month I spend 2+ hours manually auditing Terraform configs, checking for missing tags, oversized instances, and parsing Cost Explorer JSONs to find anomalies. It's repetitive, error-prone, and I know I'm leaving money on the table.

Large language models can do this — but they cost $0.01-0.03 per query, need internet, and send my infrastructure code to someone else's servers.

What if I had a specialist model that runs on a single GPU, offline, for free?

The Approach: Fine-tune Small, Think Big

For the Hugging Face Build Small Hackathon, I built FinOptix-14B — a Qwen 2.5 14B model fine-tuned specifically for:

• Auditing Terraform HCL for FinOps compliance
• Parsing AWS Cost Explorer / Anomaly Detection JSON
• Validating architecture files against governance policies
• Answering FinOps strategy questions

The constraint: ≤ 32 billion parameters. Fits on a laptop.

Training: 265 Examples, 40 Minutes, $3.50

The Dataset

I generated 265 high-quality synthetic training pairs covering:

Every output follows a structured format: findings with severity (🔴 🟡 🔵), estimated savings, and corrected code. The model learns how to think, not just what to say.

The Training

Base: Qwen/Qwen2.5-14B-Instruct
Method: QLoRA (4-bit NF4)
LoRA: r=16, alpha=32
GPU: NVIDIA A100-80GB (Modal)
Duration: 40 minutes
Cost: $3.50

That's it. No multi-GPU clusters. No days of training. One A100 for 40 minutes.

The Gotchas

It took 4 attempts to get training right:

1. ❌ Client disconnected (forgot --detach)
2. ❌ Eval split too small for sample packing
3. ❌ CUDA error from stale cache + flash attention
4. ✅ Disabled flash_attention, sample_packing, cleaned cache → success

Lesson: with small datasets (~265 examples), keep your training config simple. No fancy optimizations needed.

The Result

FinOptix-14B generates structured audit reports from raw Terraform:

Input:

resource "aws_instance" "api" {
  ami           = "ami-0c55b159"
  instance_type = "m5.4xlarge"
}

Output:

### FinOptix Audit Report

**Findings:**
1. 🔴 Over-provisioned — m5.4xlarge detected
2. 🟡 Missing governance tags
3. 🔵 Not using Graviton instances

**Estimated Impact:** ~67% cost reduction → t3.large

**Recommended Fix:**
[corrected HCL with tags and rightsized instance]

It runs on a single L4 GPU (24GB) in under 20 seconds per query. Fully offline once loaded.

Serving: The VRAM Battle

Serving a 14B model in a hackathon budget taught me things:

Key learnings:

• Use do_sample=False (greedy) with 4-bit quantization — sampling produces NaN
• Add a threading lock — concurrent model loads crash the container
• One request at a time — KV cache fills remaining VRAM

What I'd Do Differently

1. Start with 7B — 14B is at the edge of what fits on affordable GPUs. A 7B model would serve faster and cheaper.
2. More diverse outputs — My training data is good but narrow. More edge cases would help generalization.
3. Pre-quantize to GGUF — Would enable llama.cpp runtime and truly local inference on consumer hardware.

Try It

• 🔗 Live demo: huggingface.co/spaces/build-small-hackathon/finoptix14b
• 🔗 Open weights: huggingface.co/ccortezb/FinOptix-14B
• 🏕️ Hackathon: Build Small Hackathon
• **Hackaton Space FinOptix: ** build-small-hackathon/finoptix14b

Built by Carlos Cortez — AWS Community Hero, Lima, Peru. Fine-tuned on Modal, served on Hugging Face Spaces, trained on synthetic data. Total cost: $4.08.

💡 The Opportunity Cost of NOT Having It

From cost engineering: it's not just what you pay to build — it's what you lose by not building.

Conservative estimate: 2 hrs × $75/hr × 12 months = $1,800/year in engineer time saved. Plus avoided cloud waste (rightsizing + idle resources): $500-2,000/year for a small account.

ROI: $4.08 invested → $2,300+/year in value. That's a 56,000% return.

📋 But Wait — The Hidden Costs You're NOT Considering

The $4.08 model is the easy part. In the real world, adopting FinOps properly involves costs most teams ignore:

Total realistic investment for a 5-person team: $15,000 - $30,000 in Year 1.

Why this matters: A DevOps engineer or a cloud architect watching YouTube tutorials won't get you there. FinOps requires:

• Financial modeling skills (not just infra skills)
• Organizational change management (showback → chargeback)
• Data governance (billing data = business intelligence)
• AI-assisted automation (Coding Agents like Kiro + Claude are force multipliers, not toys)

The $4.08 model doesn't replace the journey — it accelerates it.

FinOptix-14B gives your team a private, local specialist that's available 24/7 during that transformation. It's the tool that makes the $15K-$30K investment actually pay off faster.

The real cost isn't $4.08. It's the $2,300/year you leave on the table every year you don't automate your FinOps — multiplied by the organizational debt of not having trained practitioners.

Final thought: After you've paid for the assessment, the training, the certifications, and the tooling — after all the hidden costs are accounted for — the $4.08 model is the cherry on top. It's the best-of-the-best optimization: a specialist that runs locally, costs nothing to operate, and compounds the value of every dollar you already invested in FinOps maturity. Totally worth it.

Check it out on : https://huggingface.co/spaces/build-small-hackathon/finoptix14b

Built by Carlos Cortez — AWS Community Hero, Lima, Peru. Fine-tuned on Modal, served on Hugging Face Spaces, trained on synthetic data. Total cost: you know now it cost more than 4.08 dollars