Where the work happened — and how to verify it.

Designed and operated an 11-agent orchestrator-worker AI research platform (~27,500 words of role-scoped agent charters) with contracted hand-off packets, few-shot routing, and separation-of-duties between generation, validation, and documentation agents — backed by ~76,500 LOC of light-dependency Python.

Built a 31-gate statistical evaluation harness (G1–G31): block-bootstrap CIs, random-timing nulls, walk-forward, 5-era stability, plus a multiple-testing layer (Deflated Sharpe, CSCV-based PBO, Minimum Backtest Length) — implemented scipy-free in numpy and used to validate both LLM outputs and systematic-trading strategies.

Architected a tiered model-routing policy (Opus = judgment / Sonnet = assembly / Haiku = mechanical / Fable = hardest autonomous) with per-dispatch token budgeting — the cost layer that keeps a multi-agent system (measured at ~15× single-agent token cost) economical to run continuously — plus structured-output contracts gated by a mechanical validator (must exit 0).

Researched and validated systematic trading strategies across 5 asset classes (equity-index, crypto, energy, metals, agriculture) on the desk’s statistical-filter platform; shipped candidates through paper-shadow and live forward out-of-sample testing under pre-registration and frozen-spec evaluation.

Built point-in-time, look-ahead-disciplined data pipelines from scratch on free public sources; implemented byte-range subsetting, completeness guards, gap logging, and idempotent incremental pulls to produce reproducible research-grade datasets.

Caught and documented false positives as enforced methodology — banking each into the desk’s reusable research-integrity playbook — and designed an automated monthly forward-out-of-sample monitoring fleet (scheduled data pull → S3 sync → frozen-spec evaluation → ledger) that produces un-gameable live performance evidence.

Built a 7-agent venture-incubation pipeline (5 judgment-tier + 2 mechanical) with 10 lifecycle skills; produced 31 decision-grade artifacts and engineered an anti-self-approval governance pattern (the agent that proposes never approves).

Encoded Philippine tax/regulatory rules (DTI/SEC/BIR/LGU, ₱3M VAT threshold, 8%-flat vs graduated election) directly into agent and skill prompts, so the research output is jurisdiction-aware at the prompt level — not bolted on at the end.

Shipped 31 decision-grade artifacts (charters, briefs, ledger entries) under the same separation-of-duties principle used in production research desks.

Built an 8-agent content-production pipeline (topic-scout → researcher → drafter → editor → producer → art-director → chart-maker → exporter) and a 290-line AI-output (slop) evaluator scoring drafts on 13 literature-grounded metrics; drove a real draft from HEAVY (index 81) to CLEAN (index 3).

Engineered a dependency-free rendering pipeline (HTML/SVG → headless-Chrome PNG; Markdown → publish-ready PDF) so the output side of the platform runs without LLM API costs.

Demonstrated that AI-assisted editorial work can be made measurable: each draft’s slop score, iteration delta, and shipped-iteration provenance are all logged.

Conducted structured testing of trading platform onboarding flows to evaluate execution clarity, workflow logic, and system usability from a trader-first perspective.

Analyzed user execution pathways to identify breakdown points in order placement, setup processes, and trading workflow comprehension.

Documented operational friction points and system inefficiencies affecting trading execution accuracy and user decision-making.

Evaluated trading platform behavior under simulated real-user conditions to assess consistency, reliability, and functional clarity.

Translated execution observations into structured insights to support product and trading workflow optimization.

Applied AI prompting tools (ChatGPT, Claude, Gemini) to accelerate literature review and pattern recognition while manually verifying accuracy of all outputs.

Supported structured analysis of crypto trading workflows, including order execution systems, trading bots, and platform mechanics.

Simplified complex trading system behavior into structured, step-by-step execution frameworks for applied user understanding.

Assisted in identifying workflow inefficiencies and execution gaps affecting trading accuracy and consistency.

Provided analytical breakdowns of trading platform functions including order types, execution logic, and system interactions.

Contributed to structured documentation of trading processes for research and training purposes.

Designed and shipped multi-agent LLM systems end-to-end: agent charters, eval harnesses, model-routing policy, structured-output contracts, persistent memory, and Python tooling — for content automation, venture incubation, and personal-knowledge workflows.

Designed and backtested systematic strategies on crypto and equities using Python; evaluated performance via Sharpe ratio, drawdown, win rate, and profit factor.

Read and summarized 20+ academic and practitioner papers on momentum, mean-reversion, volatility carry, statistical arbitrage, and multi-agent AI architectures into structured one-page research notes.

Built, audited, and open-sourced a portfolio of runnable AI projects: RAG scorecard, ReAct tool-calling agent with OTel traces, LLM-as-judge validated vs humans, MCP eval server, self-critiquing reflection agent, AI-slop evaluation gate.

Maintained a research journal documenting hypotheses, methodology, statistical tests, and outcomes — building a personal library of structured AI + quant knowledge.

Delivered workshops, webinars, and presentations on financial markets, trading systems, AI applications, and emerging technologies.

Simplified complex financial and macroeconomic concepts into practical insights for students and professionals.

Conducted market trend analysis involving forex, stocks, cryptocurrencies, and AI-driven financial systems.

Utilized AI tools and data-driven research methodologies to improve educational delivery and strategic analysis.

Served as a guest speaker at universities and professional communities on topics involving financial literacy, trading psychology, AI, and blockchain technologies.

Live lecture

Delivered to USeP EGE 313 (Understanding the ICC and Indigenous Group in Mindanao) for ECE 3A on the 2:30pm MW schedule, AY 2024-2025. 19-slide deck on the Manobo people of Mindanao, indigenous cultural communities (ICCs) and indigenous peoples (IPs), and the cultural-rights framework under the UNDRIP and the Philippine IPRA (RA 8371).

The recording shows the PowerPoint Edit-mode title (MACION. MANOBO. PIC. ECE3A. 230pm. MW) and the live webcam feed in the lower-right corner — provenance for both the deck and the speaker.

AI-assisted authoring pipeline

Three weeks before the lecture, the slide deck was drafted end-to-end via a ChatGPT-assisted authoring pipeline. The recording below shows the AI-engineering workflow that produced the deck — the same precursor pattern that today runs through a 31-gate evaluation harness before any artifact ships.

This is the lineage of the eval-first method: a 2024 single-agent drafting flow → the current 11-agent orchestrator + 31-gate harness. Public, NDA-clean, on-disk provenance preserved.