ML researcher post-training small open-weight models with SFT, DPO, and GRPO — building preference datasets and LLM-judge evals, and measuring what each stage improves and what it breaks. Full provenance and traceable research methodology.

Skills
Post-Training

SFT, DPO, GRPO / RLHF-style preference optimization, LoRA, knowledge distillation (on-policy, teacher–student), TRL, PyTorch, Transformers

Preference Data & Reward Signals

Preference-pair dataset construction, reward-signal design (GRPO), LLM-as-judge (blind pairwise, naturalness), NLI grounding veto, human-blind evaluation protocols

Evaluation & Rigor

Pre-registered held-out splits, bootstrap CIs, decision agreement, Holm correction, instruction-following / reasoning / abstention regression audits

Professional Experience
Experimental Intelligence⁠, Machine Learning Researcher: OpenGrad Research Lead
08/2026 – Present
  • •Ran a pre-registered SFT + DPO post-training pipeline for Qwen3.5-2B on a 173,237-record canonical corpus, with DPO on 481 preference pairs: SFT raised call F1 from 0.6264 to 0.7470 on the 1,277-item confirmatory split, mostly by cutting unneeded tool calls (the base model called tools on 62% of no-call items).
  • •Diagnosed what post-training broke, then acted on it: zero-shot GSM8K fell 67.4% → 0.0% (100% refusal on bare arithmetic), 8-shot fell 70.4% → 55.5%, IFEval strict fell 67.8% → 45.8%, and MMLU-Pro 49.0% → 37.0%; withdrew the DPO checkpoint's promotion when its +0.0078 gain proved within run-to-run noise, and reopened the study to explain the regression before trying to correct it.
  • •Quantized one model to nine GGUF levels from BF16 and selected Q6_K under a pre-registered rule, with 98% decision agreement with BF16 (25 of 1,277 flips).
    • SapinSapinAI⁠, DiaGrad Research Lead`
    05/2026 – 10/2026
  • •Built a 4.06M-pair deduplicated English–Filipino translation dataset (135.6M tokens, screened from 34 candidate corpora) absent from NLLB's 63.6M-line English–Tagalog data, and cleaned 70 of 2,033 MASSIVE-dev items overlapping NLLB before evaluation.
  • •Benchmarked 6 translation models (3B–32B) on 2,000 directional examples with chrF++ (95% paired-bootstrap CIs) and a blind LLM judge naturalness check.
  • •Removed 70 of 2,033 MASSIVE-dev items overlapping NLLB before evaluation, and wrote a shard-wise loader (peak CPU memory 27.9 GB → 0.6 GB) to evaluate Qwen3-32B on an NVIDIA GB10.
    • Projects
    Small-Mind Companion⁠, Multimodal Post-Training: SFT, DPO, and On-Policy Distillation
    08/2026 – Present
  • •Post-trained gemma-4-E2B through SFT → DPO → on-policy distillation from an E4B teacher (TRL); the distilled student won 38.1% of blind pairwise comparisons against 30.5% for the DPO-plus-memory variant.
  • •Improved safety and abstention behavior: correct abstention on 80 unanswerable probes rose from 13.75% to 71.25%, and a BM25 + embedding retrieval memory raised recall on 608 probes from 0.16% to 15.13%, reaching 18.59% after on-policy distillation.
  • •Published 35 corrections in an errata against the artifacts they contradict — including a judge-rubric bug that made an impossible 94% baseline visible.
    • SchemaForge⁠, Knowledge Distillation for Structured Extraction
    08/2026 – Present
  • •Distilled MiniCPM5-1B from Gemma 4 31B for JSON extraction across 12 schemas, training on synthetic hard examples filtered by a teacher-output validation gate.
  • •Paired deterministic rules with the distilled model, which more than doubled field F1 over rules alone (0.29 → 0.68) on a synthetic 12-schema benchmark including 3 document types excluded from training.
    • MoE Agentic Development Pipeline⁠, Coding-Agent Skills Benchmark
    08/2026 – 10/2026
  • •Designed and ran a pre-registered 4-arm coding-agent benchmark (120 autonomous runs, Holm-corrected comparisons); caught an agent gaming the harness by reading its own scorer (524,194 tokens, 44 tool calls) and sandboxed file tools in every arm.
    • SPEAKERSHIP
    Certificates
    Awards
    ASES 0 to 1 AI Startup Bootcamp Grand Winner⁠, Affiliated Stanford Entrepreneurial Students Manila
    29/06/2025
    eGovPH Hackathon 2026 Top 10 Finalists⁠, Department of Information and Communications Technology
    22/07/2026
    Education
    Bachelor of Science in Computer Science, Technological Institute of the Philippines
    08/2025 – Present

    Bachelor of Science in Computer Science with Specialization in Artificial Intelligence

    Far Eastern University Institute of Technology
    08/2024 – 08/2025