ClearSignal — Evaluation-First Fraud Explanation System

02 — Approach

Solution

The workflow below is a visual summary of the documented implementation; it is not a claim of additional system components.

1. 01

   Built the full pipeline that includes domain pretraining, LoRA fine-tuning (SFT and DPO), and vLLM serving with two hot-swapped adapters, reaching sub-second p95 latency (roughly 4–10× faster than a GPT-4o call) and cutting GPU memory ~47% by sharing one base model.

   →
2. 02

   Wrote an evaluation harness with automated hallucination checks, adversarial test cases, and an LLM judge that decided whether the model was good enough to ship, using pass/fail criteria fixed in advance so results couldn't be gamed.

   →
3. 03

   Added a router that keeps easy cases on the local model and escalates hard ones, which cut frontier-model cost about 63%, plus three safety guards that catch missing or malformed inputs before the model can respond.

03 — Engineering judgment

Key decisions

• 01

  Built the full pipeline that includes domain pretraining, LoRA fine-tuning (SFT and DPO), and vLLM serving with two hot-swapped adapters, reaching sub-second p95 latency (roughly 4–10× faster than a GPT-4o call) and cutting GPU memory ~47% by sharing one base model.

• 02

  Wrote an evaluation harness with automated hallucination checks, adversarial test cases, and an LLM judge that decided whether the model was good enough to ship, using pass/fail criteria fixed in advance so results couldn't be gamed.

• 03

  Added a router that keeps easy cases on the local model and escalates hard ones, which cut frontier-model cost about 63%, plus three safety guards that catch missing or malformed inputs before the model can respond.

04 — Evidence

Results

This project record verifies the delivered implementation, but does not report a comparable performance metric.

Product & System Views