AI#01
open sourcerag-recall
RAG service that proves its own retrieval — recall@3 = 0.886, MRR@3 = 0.805 with offline stdlib TF-IDF retriever.
- recall@3
- 0.886
- over 35 labeled Q's
- MRR@3
- 0.805
- Mean faithfulness
- 1.00
Portfolio
15 projects, two lanes.
Every project is reproducible, headline metrics are real, and the methodology is documented. Filter by lane to focus on the work most relevant to the role you're hiring for.
Lane 1 · AI Engineer
Multi-agent LLM systems & eval
6 projects spanning RAG scorecards, ReAct tool-calling, MCP servers, LLM-as-judge validation, reflection loops, and AI-output quality gates.
AI#02
open sourcetoolcall-agent
ReAct-style tool-calling agent with OTel traces, fault injection, and 100% tool/arg correctness.
- Tool / arg correctness
- 100%
- Injected faults recovered
- 6 / 6
- Loops without bound
- 0
AI#03
open sourcejudge-harness
LLM-as-judge pipeline validated against human raters — Cohen's κ = 0.58 with bootstrap CI and position-bias measured.
- Cohen's κ (vs human)
- 0.58
- moderate agreement
- Pass rate
- 0.60 ± 0.02
- 95% bootstrap CI
- Position bias
- 17%
- lower = better; flip-sensitive
AI#04
open sourceeval-mcp-server
MCP server exposing the slop-evaluation gate over Tools, Resources, and Prompts — 20/20 conformance, 100% round-trip parity.
- MCP conformance
- 20 / 20
- across all 3 primitives
- Round-trip parity
- 100%
- Primitives exposed
- 3 / 3
- Tools · Resources · Prompts
AI#05
open sourcereflect-revise
Reflection-loop agent — mean SLOP 127.5 → 14.0 across drafts, 3/4 improved, 1/4 honest no-progress halt.
- Mean SLOP score
- 127.5 → 14.0
- across drafts
- Drafts improved
- 3 / 4
- Honest no-progress halts
- 1 / 4
- shipped original
AI#06
open sourceslop-scanner
13-metric literature-grounded AI-output quality gate — drove a real draft from HEAVY (81) to CLEAN (3).
- Real-draft improvement
- 81 → 3
- HEAVY → CLEAN
- Metrics
- 13
- External dependencies
- 0
Lane 2 · Quantitative Researcher
Systematic strategies & statistical validation
9 projects across multiple-testing, cross-sectional and time-series alpha, volatility, cointegration, funding-carry, and bias control.
Quant#01
public-data reproducibleMultiple Testing & the Deflated Sharpe Ratio
Best-of-160 BTC rule: IS Sharpe 1.14 is only 1.24× the pure-noise expectation — DSR = 0.70 (fail).
- Best-of-160 in-sample Sharpe
- 1.14
- Pure-noise expectation
- 0.92
- Deflated Sharpe Ratio
- 0.70
- fails the multiple-testing adjustment
Quant#02
public-data reproducibleCross-Sectional Momentum (18 coins)
IS Sharpe 0.91 → OOS −0.03 — an honest decay; bootstrap CI straddles zero.
- In-sample Sharpe
- 0.91
- Out-of-sample Sharpe
- −0.03
- honest decay
- OOS 95% CI
- straddles 0
- cannot reject the null
Quant#03
public-data reproducibleTime-Series Momentum + Vol Targeting
Vol-targeting lifts Sharpe 0.27 → 0.39 and halves max DD (−62% → −30%).
- Sharpe (raw)
- 0.27
- Sharpe (vol-targeted)
- 0.39
- Max DD (raw → targeted)
- −62% → −30%
Quant#04
public-data reproducibleThe Variance Risk Premium (VIX vs Realized)
Implied > realized 85% of 36 yrs; predicts returns, Newey-West t = +6.5.
- Years VRP > 0
- 85%
- of 36 yrs
- Newey-West t-stat
- +6.5
- HAC-consistent
- Sample window
- 1990 – 2025
Quant#05
public-data reproduciblePairs Trading via Cointegration (BTC / ETH)
ADF −2.11 (not cointegrated), half-life 208 d — the fade loses, as the test predicts.
- ADF test statistic
- −2.11
- fail to reject unit root at 5%
- Half-life (estimated)
- 208 days
- OOS Sharpe (fade)
- ≈ 0
- as the test predicted
Quant#06
public-data reproducibleCrypto Funding-Carry
Funding +11.9% annualized premium; fade IS 1.11 → OOS −0.05 (decayed post-2023).
- Annualized funding
- +11.9%
- BTC perp, ~30-day centered mean
- Fade IS Sharpe
- 1.11
- Fade OOS Sharpe
- −0.05
- decay post-2023
Quant#07
public-data reproducibleMacro / Volatility-Regime Overlay
Vol-managed exposure: Sharpe 0.64 → 0.67, max DD −58% → −42%.
- Sharpe (base)
- 0.64
- Sharpe (regime overlay)
- 0.67
- Max DD (base → overlay)
- −58% → −42%
Quant#08
public-data reproducibleBacktest Engine + Cost Model
High-turnover signal wins gross (0.20 > 0.13) but loses net of cost — break-even 20 bps.
- Sharpe (gross)
- 0.20
- Sharpe (net, realistic)
- < 0
- loses after cost
- Break-even cost
- ~20 bps / RT
Quant#09
public-data reproducibleLook-Ahead Bias Audit (the shift test)
A leak inflates Sharpe 0.59 → 5.07; the shift test exposes it as 88% phantom.
- Clean Sharpe
- 0.59
- Leaked Sharpe
- 5.07
- Phantom % detected
- 88%
- of the inflated Sharpe is leakage
Combined-book analysis · cross-project
What 9 projects tell you about a research book
None of these projects stand alone. The signal, cost, regime, and bias-control layers are designed to compose — a single vol-targeted carry trade (03 + 06) carries different risk than a cointegration pair (05) carrying regime-aware sizing (07). The shared methodology across all nine is what makes the book interpretable to a risk manager, not just an alpha-hunter.
Layer A — Signal
Alpha sources
- Cross-sectional: project 02 (failed OOS, as the test predicted)
- Time-series: projects 03 (vol-targeted), 04 (VRP)
- Carry / funding: project 06 (OOS Sharpe decayed post-2023)
- Cointegration / stat-arb: project 05 (rejected by ADF)
Layer B — Risk
Risk-management overlays
- Vol targeting (20-day): project 03 (Sharpe 0.27 → 0.39)
- Regime overlay (smoothed VIX): project 07 (DD −58% → −42%)
- Cost-aware sizing: project 08 (break-even turnaround scan)
- Look-ahead audit: project 09 (one-line shift test, every book run)
Layer C — Validation
Statistical gating
- Multiple-testing: project 01 (Deflated Sharpe, 160-trial null)
- Pre-registration: project 06 (OOS window locked before data touched)
- HAC inference: project 04 (Newey-West t = +6.5)
- Honest nulls: projects 02 + 05 (rejected, shipped as rejections)
The book is not nine winning trades. It is one pipeline with three orthogonal layers and a layered validation discipline that catches the failure modes that eat naive books alive. A risk manager reading these nine projects sees the same engineering discipline applied end-to-end — that is what is hireable.