solutions/ · case studies
I do solutions.
Each entry below is a real engagement or self-directed project: the problem, the approach, the evidence, the outcome, and the proof. Every claim links to a falsifiable test. None of it relies on proprietary data sources.
decision=argmaxa∈AE[U(a)∣evidence]
PROMISE
Every solution ships with at least one falsifiable claim and a proof pointer.
professional certsc(t)
102
11-month arc
LOC light-dep Python∣L∣
76.5k
numpy/pandas/boto3
eval gatesg
31
G1–G31 statistical
Lane · AI Engineering
Multi-agent systems that pass validation.
11-Agent Eval-First Research Platform
19V Capital (closed past contract, 03/2026 – 06/2026)
Problem
A small systematic-trading desk needed an AI workflow that could keep pace with the research pipeline
*without* shipping hallucinated or unverifiable analysis. The naive path — a single agent with one
prompt — produced plausible but unfalsifiable outputs. The desk needed something closer to a research
organization than a chatbot.
Approach
Built an **orchestrator + worker topology** with 11 agents, ~27,500 words of role-scoped charters, and
contracted hand-off packets between generation, validation, and documentation roles. Added a 31-gate
statistical evaluation harness (G1–G31) that ran the same validation stack on LLM outputs and
on the desk's systematic trading strategies — so the *same* notion of "evidence" applied across
both. Tiered model-routing (Opus = judgment, Sonnet = assembly, Haiku = mechanical) kept the
~15× token multiplier economical.
Evidence
- **~27,500 words** of role-scoped agent charters
- **31-gate** statistical evaluation harness (G1–G31)
- **~76,500 LOC** light-dependency Python
- **5 asset classes** researched (equity-index, crypto, energy, metals, agriculture)
- **~15×** token cost vs single-agent, offset by tiered routing
Outcome
Caught and documented false positives as enforced methodology (each banked into the desk's research-integrity
playbook). Shipped an automated monthly forward-OOS monitoring fleet — scheduled data pull → S3 sync →
frozen-spec evaluation → ledger — that produced un-gameable live performance evidence.
Proof
- Contract disclosed under NDA — PM of record (publicly attributable): Evan Ferioli
- Eval harness methodology (G1–G31, deflated Sharpe, block-bootstrap CIs, walk-forward) carried forward into the public methodology page
- All deliverables NDA-safe; no proprietary data sources referenced
7-Agent Venture Incubation Pipeline
Macion Ventures
Problem
An operator-led venture incubation arm needed a repeatable way to triage, brief, and ledger new
business ideas — without one human bottleneck, and without the agent that proposes a decision
being the same one that approves it.
Approach
Built a **7-agent pipeline** (5 judgment-tier + 2 mechanical) with 10 lifecycle skills and 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.
Evidence
- **31 decision-grade artifacts** (charters, briefs, ledger entries) shipped
- **Anti-self-approval** governance pattern documented as a reusable convention
- **Jurisdiction-aware prompting** for PH tax/regulatory rules
Outcome
Produced a clean separation-of-duties trail — every proposal had a corresponding validator hand-off,
the validator never originated the proposal, and every decision was ledgered for later audit.
The same separation-of-duties principle was reused at 19V for LLM research validation.
Proof
- 31 artifacts on disk (charters, briefs, ledgers)
- Reuse pattern documented in 19V engagement methodology
8-Agent Editorial Production Pipeline (SLOP ↓ 96%)
Editorial AI / Content Automation
Problem
A high-volume editorial workflow was generating content with a measurable "slop index" — generic,
templated, easily-detected text. Quality gate was after-the-fact and manual. Production scaled
faster than the editorial team could review.
Approach
Built an **8-agent content production pipeline** with a **SLOP-scanner** (proprietary eval gate)
woven into the pipeline as a mechanical validator. The scanner measured the proportion of stock
phrases, hedge words, and un-statistical copy against a baseline corpus, and gated publication.
Agents were chartered to **rewrite before publishing**, not to publish first and edit later.
Evidence
- **SLOP index dropped from 81 → 3** on the working corpus
- **Mechanical validator** enforced the gate (exit-0 contract)
- **Rewrite-before-publish** semantics in the agent charter
Outcome
Production volume held; editorial-review labor fell (the gate did the rejection). The
SLOP-scanner is one of the OSS projects in the AI portfolio (separate repo).
Proof
- Scorecard JSON for the SLOP scanner is OSS
- Pipeline charter documented in editorial-ai engagement
Eval MCP Server — 31 Gates as First-Class Tools
Self-directed / OSS
Problem
LLM eval harnesses live in code or in spreadsheets — neither is a clean integration target for
multi-agent pipelines. A multi-agent system needs the eval gates exposed as **tools**, not as
Python imports.
Approach
Built an **MCP server** that exposes the 31-gate statistical evaluation harness as discoverable
tools. Each gate has a typed contract (input schema, output schema, exit codes). An agent can
dispatch a strategy candidate through the gate stack via MCP without owning the implementation.
**The mechanical validator enforces exit-0** before downstream agents consume the result.
Evidence
- **MCP-compliant** server (Claude Agent SDK / Cursor / etc. can connect)
- **31 typed tools** (one per gate)
- **Exit-0 contract** enforced by mechanical validator
- **No agent can bypass** the gate stack when calling downstream
Outcome
The eval harness becomes the **single source of truth** for what counts as "passed validation" —
whether the input is an LLM output or a systematic strategy. Reusable across projects.
Proof
- Open-source repo (eval-mcp-server) with scorecard
- Documented in AI portfolio README
9-Project Public-Data Quant Research Library
Self-directed / Portfolio (public-data reproducible)
Problem
Most online quant research demos are not reproducible: closed datasets, undisclosed parameters,
un-reported multiple-testing bias, and no OOS discipline. A hiring-grade research portfolio
needs every one of those addressed explicitly.
Approach
Built **9 reproducible research projects** on free public data — multiple-testing (Deflated Sharpe),
cross-sectional & time-series alpha, volatility carry, cointegration, funding-carry, regime
overlays, transaction-cost realism, and look-ahead-bias audits. Each project shipped with a
locked OOS window, block-bootstrap CIs, frozen-spec evaluation, and methodology stated up-front
(not bolted on).
Evidence
- **9 projects**, each with methodology declared before results
- **Multiple-testing discipline** — Deflated Sharpe, CSCV-based PBO, MinBTL
- **Locked OOS windows** + **block-bootstrap CIs** on every project
- **Look-ahead-bias audits** as a dedicated project
Outcome
Library is recruiter-readable in <10 minutes per project; methodology gates are stated *before*
numbers so a hiring manager can see the rigor upfront.
Proof
- Each project has a memo + figure (public)
- All projects on free public data — fully reproducible
Deflated Sharpe Ratio as a Built-in Pipeline Gate
19V Capital (closed past contract, 03/2026 – 06/2026)
Problem
Naive Sharpe ratios ignore the search effort — if you try 100 variants of an idea, the
best-looking one will overstate the true edge. The desk needed this multiple-testing
correction built into the validation pipeline, not stapled on at the end.
Approach
Implemented **Deflated Sharpe Ratio (DSR)** with a scipy-free numpy implementation, plus
**CSCV-based Probability of Backtest Overfit (PBO)** and **Minimum Backtest Length (MinBTL)** —
the three canonical multiple-testing corrections. All three run as gates G-23 → G-25 in the
31-gate harness, every strategy candidate.
Evidence
- **scipy-free** numpy implementation (deploys anywhere)
- **3 canonical corrections** in one gate stack (DSR, PBO, MinBTL)
- **All gates share the same contract** as the LLM-eval gates
Outcome
No strategy ships through the pipeline without surviving the multiple-testing correction layer.
This gate is reusable: any new systematic strategy passes through the same validation stack.
Proof
- Methodology referenced in 19V public statements
- Implementation available in public quant library
Crypto Statistical-Arbitrage Pipeline with Funding-Carry
Self-directed / Ledger51 era (public-data)
Problem
Crypto perps run funding payments every 8h. A naive long-short book ignores funding carry and
bleeds slowly when the spread is inverted. A working stat-arb needs the funding cost added back
to P&L *before* sizing.
Approach
Built a **crypto stat-arb pipeline** that integrates funding carry into the cost model, pairs it
with a cointegration gate (Engle–Granger + Johansen + half-life band), and gates trades through
the same 31-gate harness as systematic equity strategies. Locked OOS window; block-bootstrap CIs;
random-timing nulls.
Evidence
- **Funding-carry** integrated into P&L (not an afterthought)
- **Cointegration + half-life** gate before trade signal
- **Same 31-gate harness** as equity strategies
- **Public-data reproducible** (Binance/Bybit free data)
Outcome
Pipeline gates every candidate through the standard gate stack. Funding carry is part of the
cost model, not a P&L surprise. Research-ready for hedge-fund desks that already understand
carry-aware systematic trading.
Proof
- Memo + figure in public quant portfolio
- Methodology declared in project README
Transaction-Cost-Aware Backtest Engine
Self-directed / Portfolio
Problem
Backtests that ignore transaction costs, slippage, and latency are the most common source of
overfit research. A backtest engine needs all three treated as first-class inputs, not as
after-the-fact deductions.
Approach
Built a **transaction-cost-aware backtest engine** in numpy (no scipy dependency) with realistic
spread + slippage + latency models per asset class. The engine treats **capacity** as a
constrained variable — strategies report the AUM at which they would still survive costs.
Walk-forward and 5-era stability are enforced.
Evidence
- **Spread + slippage + latency** modeled per asset class
- **Capacity constraint** enforced as a reportable dimension
- **numpy-only** (deploys anywhere)
- **Walk-forward + 5-era stability** gates on every strategy
Outcome
Strategies ship with a **capacity-aware** backtest, not a fantasy one. A hiring manager can
read a project's "survives costs at $X AUM" line and immediately trust the number.
Proof
- Memo + figure in public quant portfolio
- Methodology documented in project README
Look-Ahead-Bias Audit Suite
Self-directed / Portfolio
Problem
Look-ahead bias is silent: a backtest that uses future data looks great until you deploy it.
A serious research portfolio needs an **explicit audit suite** — a checklist of failure modes
with mechanical tests.
Approach
Built a **look-ahead-bias audit suite** with mechanical tests for the most common failure modes:
point-in-time dataset verification (using public timestamps, not internal close-times), survivorship
bias checks, rebalance-time vs. signal-time consistency, and frozen-spec evaluation. Each test is
a gate that has to pass before a candidate is allowed to report a number.
Evidence
- **Point-in-time** dataset verification
- **Survivorship bias** checks
- **Rebalance/signal timing** consistency tests
- **Frozen-spec evaluation** prevents post-hoc tuning
Outcome
The audit suite is run on every strategy candidate. A pass means the candidate's reported
numbers are point-in-time consistent and have not been tuned to the test set.
Proof
- Public OSS audit suite + memo
- Documented in quant portfolio
Public Finance Curriculum (CTA-Track, Self-Directed)
Public — finance students / early-career analysts
Problem
The Philippines has limited access to rigorous CTA-grade technical analysis training. Most
curriculum is either imported (US/UK, expensive) or shallow (TA-by-rote). There's a gap for
affordable, rigorous, evidence-based technical analysis education.
Approach
Designed and taught a self-directed **CTA-track public curriculum** aligned with the Society
of Technical Analysts (STA) Tier-1 syllabus. Built **animated explainer videos** for hard
concepts (Black-Scholes intuition, Monte Carlo via reproducible notebooks, regime overlays,
multiple-testing discipline). Open-sourced notebooks so students can re-run everything.
Evidence
- **STA Tier-1 CTA** certified (program completed Dec 2025)
- **Public animated explainer videos**
- **Open-source notebooks** (reproducible)
- **University guest lectures** at PSHS-SMC and USeP
Outcome
Curriculum reached students who would otherwise not have had access to CTA-grade training.
Notebooks and videos are public — anyone can re-run the analysis end-to-end.
Proof
- Public video library (URL on contact page)
- University guest-lecture invitations (PSHS-SMC, USeP)
- STA Tier-1 CTA certificate
Want a deep-dive on any of these?
Email me the JD or scope and I'll send back a tailored solution sketch within 24 hours.
