Data scientist who builds data pipelines, document extraction and RAG systems in Python and SQL, and backs every result with reproducible evaluation (held-out sets, confidence intervals, regression gates).

Skills
Data & Pipelines

Python, SQL, PostgreSQL, Pandas, NumPy, Data Cleaning, PDF Extraction (GROBID, PyMuPDF, Docling)

NLP, LLMs & RAG

Entity Extraction, RAG, Hybrid Search, Reranking, LangGraph, Tool Calling, PyTorch, Transformers, LoRA/SFT/DPO

Evaluation & MLOps

Recall@k/MRR, LLM-as-Judge, Confidence Intervals, Git, GitHub Actions CI/CD, Docker, AWS EC2, vLLM

Professional Experience
Experimental Intelligence⁠, Co-Founder & Machine Learning Researcher
08/2026 – Present
  • •Built a reproducible data pipeline that cleaned 199,260 rows from 4 pinned public datasets into 181,433 deduplicated, validated examples, with all 29 outputs byte-identical across rebuilds.
  • • Built a GPU-free re-scoring framework (48,840 per-example records) that caught the promoted model refusing all 1,319 zero-shot GSM8K questions its gate had passed.
  • •Fine-tuned Qwen3.5-2B for tool calling (SFT + DPO), lifting call F1 0.6264 → 0.7548 on a 1,277-example held-out set, with the decision rule fixed before results.
  • •Audited ~990 published claims against committed results and corrected 100, codified as 17 guardrail rules with a CI-checked claim ledger.
    • Agylion⁠, AI Engineer - RepReady Technical Head
    05/2026 – Present
  • •Built multi-tenant hybrid RAG on Postgres (pgvector + full-text search, RRF fusion), with an optional reranker that falls back safely and row-level-security tenant isolation.
  • •Built a post-call LLM processing pipeline: retried, deduplicated job queues with retry-safe checkpoints and a recovery sweep for stuck jobs. Covered by 339 tests and GitHub Actions CI.
  • •Hardened LLM pipelines with 5-level prompt-injection defence, PII/secret redaction and audit logs; grading runs on retried, per-session-locked queues with a deterministic fallback.
  • •Protected client data with PII/secret redaction and 5-level prompt-injection defence; wrote a 20-page white paper projecting 4.7× lower voice cost per call ($1.05 → $0.224).
    • Projects
    OpenPapers⁠, Provenance-first research MCP server + multi-agent pipeline
    08/2026 – Present
  • •Built ingestion from 6 public APIs with PDF extraction (GROBID vs PyMuPDF), scoring title F1 0.976 and abstract F1 0.957 on a 26-paper gold set.
  • •Resolved paper identities across sources with 1.0 accuracy on 119 alias cases (0 false merges), and reached Recall@10 0.856 / MRR 0.773 on 44 retrieval queries.
  • •Built a LangGraph multi-agent pipeline with an NLI grounding check, lifting cited-sentence support 81.6% → 99.4% on a 5-question evaluation for $0.24; CI with 274 tests.
    • Small-Mind Companion⁠, A Multimodal Post-Training Project with External Memory and Post-Training
    08/2026 – Present
  • •Extracted structured memory facts from conversations with an LLM plus verbatim-span validation (503 of 505 claims accepted), stored for BM25 + embedding retrieval.
  • •Tuned retrieval with a k-sweep, lifting answer accuracy 0.16% → 15.13% on 608 probes; the full fine-tuned stack raised abstention on 80 unanswerable probes 13.75% → 71.25%.
    • Certificates
    Awards
    eGovPH Hackathon 2026 Top 10 Finalists⁠, Department of Information and Communications Technology
    22/07/2026
  • •Placed top 10 at the eGovPH Hackathon 2026 by leveraging GabayMed, integrating 9 native eGov APIs to slash medical assistance processing from 14 days to real-time across multiple government agencies and political agencies for 100% auditable impact.
    • ASES 0 to 1 AI Startup Bootcamp Grand Winner⁠, Affiliated Stanford Entrepreneurial Students Manila
    29/06/2025
  • •Won ASES 0 to 1 by leveraging PocketPatient, backed by thorough market research and a clear strategy targeting an exclusive niche for scalable impact and profitability.
  • •Placed 1st runner-up at GrowthCon 2026 Killer Pitch by leveraging RepReady under Agylion, a sales activation agent that utilizes a 3-step AI simulation framework to target the 7-to-9-month enterprise ramp time and eliminate revenue leaks.
    • 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