Software Engineer, LLM Performance & Evaluation

FuriosaAI · Software · Seoul
공고 번호 4006955201

LLM 추론 속도와 모델 정확도를 재현 가능한 평가·회귀 검사로 추적.

학위 제한 미명시 · 추가 검토

확인할 점: TTFT·토큰 지연 분포·NPU/기준 구현 수치 비교와 반복 성능 benchmark는 보강 필요.

영문 이력서 PDF ↗영문 이력서 Overleaf ZIP ↓

Cover Letter PDF ↗Cover Letter Overleaf ZIP ↓

이 공고의 지원서 열기 ↗

Career Summary

필수 Career Summary 칸에 붙여 넣기.

I am an ML engineer and AI Researcher at ISPARK R&D Center responsible for TieMing's core AI features, from model training to inference APIs and deployment. I bring a multimodal research background—five publications, four as first author—and turn experimental results into working products. I evaluate model changes against task results, with explicit baselines, inputs, and scoring conditions.

I improved normalized database-result agreement from 26.6% to 76.6% on a 64-question internal CRM development benchmark through teacher-student training and LoRA. I also verified and documented the schema input required to reproduce the model's behavior in serving.

I developed business-card extraction that reached 94% accuracy across seven core fields. I applied teacher-student knowledge distillation and LoRA fine-tuning; separate difficult-case evaluation improved from 30.2% to 65.3% through further training and output correction.

I bring experience using evaluation to guide model improvements, including Text-to-SQL adaptation and business-card extraction. I define success through the result a task requires and investigate the failure patterns behind the aggregate score. At FuriosaAI, I can contribute reproducible model workloads and reference results, then trace quality or performance regressions to specific input and implementation changes. My training and evaluation experience lets me connect measurements to the next engineering decision.

Cover Letter

PDF 첨부 또는 Enter manually에 붙여 넣기.

Dear FuriosaAI Hiring Team,

I am applying for the Software Engineer, LLM Performance & Evaluation position (Job ID 4006955201). I bring experience using evaluation to guide model improvements, including Text-to-SQL adaptation and business-card extraction. I define success through the result a task requires and investigate the failure patterns behind the aggregate score. Before ISPARK, I served one year as an information-systems NCO in the Republic of Korea Air Force (Dec 2021–Dec 2022), maintaining Linux servers through shell scripting and log analysis.

I improved a four-billion-parameter Text-to-SQL model from 26.6% to 76.6% normalized database-result agreement on a 64-question internal CRM development benchmark, using teacher-student training and rank-16 LoRA. The model matched 49 of 64 reference results, compared with 17 for the base model; scoring ignores row order and normalizes numbers and dates. I also verified the serving interface: missing schema input caused invented table names despite unchanged weights. I documented the required schema inputs alongside the adapter so the trained behavior could be reproduced in serving.

I developed business-card information extraction used in TieMing, reaching 94% accuracy across seven core fields in a workflow covering approximately 5,000 cards. To address correct text being assigned to the wrong fields, I trained a smaller student model through knowledge distillation and low-rank adaptation (LoRA). I separately targeted difficult cases: follow-up training and selective regeneration of malformed outputs raised field-value accuracy from 30.2% to 65.3%. The work involved choosing the training method, diagnosing output errors, and implementing the changes in the extraction workflow.

At FuriosaAI, I can contribute reproducible model workloads and reference results, then trace quality or performance regressions to specific input and implementation changes. My training and evaluation experience lets me connect measurements to the next engineering decision.

Sincerely,
Jeonghun Kim