← INDEX/WORK/FAHE
№ 02 — PROPRIETARY LLM ENGINE
The GM who never sleeps — a fine-tuned model that answers like an operator, not a chatbot.
THESIS
Frequentor’s venues needed an AI that reads a live data snapshot — occupancy, revenue, lapsed customers, staffing, food cost — and returns one decisive, GM-grade action. No hedging, no invented numbers, no pretending to see data the POS doesn’t expose. Commodity API models couldn’t hold that persona reliably, and per-query economics at venue scale demanded a small local model.
So I built FAHE — the Frequentor Agentic Hospitality Engine — the hard way: my own datasets, my own fine-tunes, my own eval harness, my own serving stack. Over 40 model versions across two base architectures, and an ending most ML teams would recognize: the winning system wasn’t the biggest model, it was the best-orchestrated one.
THE BUILD
Mistral-7B-Instruct LoRA/QLoRA via Unsloth (rank 16, RSLoRA, all 7 target modules, 4-bit NF4) on a single RTX 5070 Ti. Hand-built datasets grew from 6,841 to 12,441 instruction pairs, with per-version “surgical” datasets generated by dedicated scripts.
QLoRA on Lambda H100s, managed fine-tunes on Azure AI Foundry, then adapter merge → S3 → AWS Bedrock custom-model import. The 70B raw model lost to the orchestrated 7B on the battery — and I kept the receipts.
Merge → BF16 GGUF via llama.cpp → Q4_K_M (4.07 GB for the 7B). Served through LM Studio behind a Node gateway with agentic tool routing to five live AppSync/Lambda tools, exposed remotely over Tailscale + Cloudflare tunnels.
41 scenarios × 10 real-venue profiles, plus a 120-test conversational battery. Category scoring across diagnostics, revenue, anti-hallucination, counterfactuals, and time pressure — with named failure patterns and token-level forensics for every regression.
Instead of training version 45, I built an inference-layer sanitizer — negation-aware output repair, decisiveness rewrites, meta-leak stripping — that lifted the frozen v11 weights from 94% to 404/410 (99%) with zero retraining. Knowing when to stop training is the skill.
A Python/SQLite venue simulator (101 MB of synthetic Toast-style POS data) where each venue gets a different POS provider profile and a visibility engine masks the fields that provider can’t see — so the model learns to say “I don’t have that data.”
ElevenLabs streaming TTS conductor with barge-in, SSML break injection, and number normalization — about 593 ms to first audio. 18/18 end-to-end voice tests passing.
Documented and fixed TorchInductor/Triton crashes on sm_120, fused cross-entropy OOMs, and stale compile-cache corruption — the unglamorous 20% that makes local training on brand-new silicon actually work.
SCREENS
SIGNALS
STACK
WHAT IT PROVES
The full ML lifecycle, solo: dataset design → fine-tuning → quantization → serving → evaluation → orchestration. And the judgment layer on top: measuring everything, freezing the weights when the data said so, and getting the last five points from engineering instead of compute.