Emmanuel Adutwum

★ FLAGSHIP C++20 / SYSTEMS

tapebench — C++20 Backtesting Engine

3.1x Faster Dispatch3.8x Parallel SpeedupTSan-Clean

Event-driven backtesting engine with a structurally-enforced no-look-ahead execution model — a strategy's signal from bar N's close can only fill at bar N+1's open, by construction, not convention. Zero-overhead strategy dispatch via CRTP (measured ~3.1x faster than virtual dispatch, isolated so the comparison is fair), a zero-copy memory-mapped tick pipeline (verified 0 heap allocations, not assumed), average-cost position accounting handling the tricky reversal-fill case, and a hand-built thread pool for parallel parameter-grid backtesting (~3.8x wall-clock speedup, proven byte-identical to sequential — not just usually correct). 74 tests, sanitizer-clean CI across gcc/clang plus a dedicated ThreadSanitizer job, real captured benchmark numbers, and a DESIGN.md documenting every real bug the process caught.

C++20CMakeGoogleTestGoogle BenchmarkGitHub Actions

NEW C++ / ML ENGINE

Neural Net Engine in C++

97.8% MNISTNo ML LibsAdam Optimizer

Full neural network library built from scratch in C++17 — zero ML framework dependencies. Custom matrix engine, backpropagation, Adam/SGD optimizers, BatchNorm, Dropout. Achieves 97.8% on MNIST. SIMD-optimized matrix multiply benchmarks included.

C++17CMakeEigenGoogle TestMNIST

NEW TRANSFORMER / GPT

GPT from Scratch (PyTorch)

Pure PyTorchNo HuggingFaceShakespeare

Complete GPT architecture from scratch — multi-head self-attention, causal masking, positional encoding, layer norm, residual connections. Character-level tokenizer. Cosine LR schedule with warmup. Top-k and nucleus sampling.

PythonPyTorchTransformersNLPLM

NEW CV / C++ INFERENCE

ResNet CIFAR-10 + C++ Inference

92.4% Acc.6x C++ SpeedupONNX Export

Custom ResNet-18 trained on CIFAR-10 in PyTorch (92.4% accuracy). Exported to ONNX and loaded into a C++ inference engine using ONNX Runtime. Benchmark shows 6x latency improvement over Python baseline. Docker deployment included.

PythonC++PyTorchONNX RuntimeOpenCV

NEW AUTONOMOUS DRIVING

Autonomous BEV Perception

Lift-Splat-ShootCenterPoint HeadKITTI 3D mAP

Production-grade Bird's-Eye-View 3D object detection pipeline. Implements Lift-Splat-Shoot (ECCV 2020) with ResNet-50 + FPN backbone and anchor-free CenterPoint detection head. Gaussian Focal Loss, mixed precision training, KITTI evaluation.

PythonPyTorch3D DetectionBEVKITTIFPN

NEW 3D RECONSTRUCTION

Neural Radiance Fields — From Scratch

31+ dB PSNRMip-NeRF IPEMarching Cubes

Research-faithful NeRF (Mildenhall et al., ECCV 2020) in pure PyTorch — no NeRF libraries. Positional encoding, hierarchical coarse-fine sampling, volume rendering integral. Achieves 31+ dB PSNR on NeRF-Blender matching the paper.

PythonPyTorchNeRF3D VisionVolume RenderingMip-NeRF

ML / REGRESSION

Optimised SVR for California Housing Price Prediction

R² 0.723 SVRRandomizedSearchCV10 Models

Optimised SVR-RBF achieving R² 0.723 on California Housing — a +0.123 absolute gain over the previously reported SVR baseline (Preethi et al., 2025). 4-stage ablation study shows scaling alone contributes +0.744 in R². Leakage-safe scikit-learn Pipeline with 10 domain-engineered features and ensemble feature selection. Submitted to Springer Nature.

PythonScikit-learnSVRAblation StudyFeature Engineering

ML / SPORTS ANALYTICS

Soccer Match Outcome Predictor

XGBoostELO RatingsForm Features

Machine learning model predicting soccer match outcomes using team ELO ratings, recent form, home advantage, head-to-head records, and player availability features. Ensemble of XGBoost and Logistic Regression with calibrated probabilities.

PythonXGBoostPandasScikit-learnELO

NEW AIRFLOW / DATA ENG

Airflow ETL Pipeline

Daily ScheduleRetry LogicData Validation

Apache Airflow DAG ingesting live crypto market data from CoinGecko. Pandas transformations (volatility, market dominance, trend features), SQLite upsert storage, automated data quality checks, and failure alerting.

Apache AirflowPythonPandasSQLitePytest

NEW INFOSEC / DEVOPS

Secure FastAPI + CI/CD Pipeline

9 TestsDockerGitHub Actions

Security event management API — API key auth middleware, Pydantic IP regex validation, per-IP rate limiting. GitHub Actions CI runs pytest + flake8 on every push, builds Docker image, and smoke-tests the container on merge to main.

FastAPIDockerGitHub ActionsPython

NEW SPARK / DELTA LAKE

PySpark Pipeline + Delta Lake

1M+ Records2–4x SpeedupPartitioned Parquet

PySpark ETL on 1M+ taxi records — schema-enforced reads, window function analytics, caching benchmarks. Delta Lake module: ACID upserts via merge, time travel queries. Partitioned Parquet output with ~60% faster downstream reads.

Apache SparkDelta LakePySparkParquet

NEW JAVA / SPRING BOOT

Spring Boot Security Alerts API

Java 17JUnit 5CI/CD

Production Java REST API — custom Jakarta servlet filter for API key auth, Bean Validation with IP regex, Spring Data JPA + H2. JUnit 5 MockMvc integration tests. GitHub Actions CI builds, tests, and smoke-tests the JAR on every push.

Java 17Spring Boot 3JUnit 5Maven

NEW AI AGENTS / FINCRIME

FinSentinel — Autonomous Financial Crime AI Platform

5-Agent Pipeline~1,500x Faster<$0.02/Run

Solo-built 5-agent AI pipeline (Gemini 2.5 Flash) automating fraud/AML investigations end-to-end — generating FinCEN SARs in under 3 minutes for under $0.02/run vs. a 3-day manual baseline. FastAPI + WebSocket streaming backend, MongoDB Atlas $vectorSearch (3072-dim embeddings), React/TypeScript frontend, Docker + Google Cloud Run deploy, modeled BSA/FINRA/OFAC/EU AI Act compliance.

Gemini 2.5FastAPIMongoDB AtlasReactTypeScriptDocker

QUANT / LIVE

Live Algorithmic Trading System

Sharpe 2.109 Strategies22 Instruments

Production trading bot on AWS EC2 — HMM regime detection (4-state), 29-feature ML signal engine, fractional Kelly sizing, automated strategy grading across forex, crypto, commodities, and indices.

PythonHMMAWS EC2OANDABinance

NEW QUANT / MOBILE

QuantLab Pro

Shipped to Google Play6 Models, Zero Libs

Cross-platform (Android/iOS/Web) quant finance app built and shipped to the Google Play Store as sole developer — React Native + TypeScript with a fully client-side compute engine, 24-tester beta before release. Six financial models implemented from scratch, zero external math libraries: Black-Scholes + full Greeks, Newton-Raphson implied vol, Avellaneda-Stoikov market-making with Monte Carlo P&L, GARCH(1,1), Hurst exponent, and Ornstein-Uhlenbeck.

TypeScriptReact NativeExpoEAS BuildBlack-Scholes

★ FLAGSHIP QUANT / FULLSTACK

Ghana Cocoa Derivatives Platform

250+ Active Users10K MC Paths, <2sSelf-Directed

Self-directed, full-stack derivatives pricing platform, live in production for Ghanaian cocoa cooperatives since Sep 2024 — not a classroom project. Extends Black-Scholes with weather-adjusted stochastic volatility (weather explains 38% of price variance in the fitted ensemble), a Monte Carlo engine (10K+ paths in under 2s), an ensemble ML layer (XGBoost, Random Forest, SVR), and an NLP assistant for non-technical users. Grew to 250+ active users organically.

Black-ScholesMonte CarloXGBoostFlask

★ FLAGSHIP QUANT / HFT

HFT Market Making Simulator

6x Higher P&L70% Lower Inventory Risk31 pytest Tests

Price-time-priority limit order book with heap-based matching (O(log n), lazy-deletion cancellation) and a synthetic market combining Brownian-motion price dynamics with Poisson order-arrival flow. Avellaneda-Stoikov optimal quoting — closed-form reservation price, online volatility estimation, inventory-based skewing — benchmarked against a naive fixed-spread baseline across 20,000-tick backtests, 5 seeds: 6x higher P&L, 70% lower inventory risk, consistent in direction on every seed. (Sharpe came out at 94.8 vs. 4.3 in this backtest — that's a short-horizon annualization artifact, explained honestly in the README, not a claim about live performance; the 6x P&L / lower-inventory-risk numbers are the ones that generalize.)

PythonNumPyPandasMatplotlibpytestAvellaneda-Stoikov

NEW DISTRIBUTED SYSTEMS

Distributed Trading Engine Benchmarking Platform

1,000+ BotsSub-20ms p99eBPF Probes

Distributed platform (4 microservices over gRPC + Redpanda) stress-testing trading engines with 1,000+ concurrent bots across 5 archetypes, streaming live scores to a real-time leaderboard — sub-20ms p99 across 62 commits and 18+ PRs. Kernel-level eBPF TC hook prober (Cilium) capturing nanosecond RTTs; full IaC (Terraform + Helm on AWS EKS) with one-command deploy; chaos injection via tc netem + cgroup v2; GitHub Actions CI/CD with cosign-signed images.

GoKubernetesRedpandaeBPFgRPCTimescaleDBTerraform

QUANT / DERIVATIVES

Black-Scholes Delta Hedging Engine

Full GreeksGamma P&LIV Surface

Complete options pricing and delta-hedging engine — Black-Scholes analytical solution, full Greeks (Delta, Gamma, Theta, Vega, Rho), implied volatility surface construction via Newton-Raphson IV solver, daily P&L attribution, and replication portfolio tracking.

PythonBlack-ScholesNumPySciPyGreeks

NEW OCAML / SYSTEMS

OCaml Limit Order Book

Price-Time PriorityMatching Engine10 Tests Passing

Functional limit order book engine in pure OCaml — price-time priority matching with BidMap/AskMap balanced BSTs using custom comparators (highest bid first, lowest ask first). Supports limit orders, market orders, partial fills, multi-level sweeps, and cancellations. Interactive REPL with live depth display.

OCamlDuneAlcotestFunctionalMap.Make

NEW OCAML / QUANT

OCaml Probability & Options Pricing

Monte CarloBlack-Scholes12 Tests Passing

OCaml library of probability puzzles and quantitative finance models — every puzzle returns (analytical, MC) pairs that converge to <0.5%. Prices options via Black-Scholes, CRR binomial trees (500 steps), and Monte Carlo (100k paths). Full Greeks via finite differences, implied vol via Newton-Raphson, and Avellaneda-Stoikov optimal market-making vs naive fixed-spread.

OCamlBlack-ScholesMonte CarloGreeksAvellaneda-Stoikov

★ FLAGSHIP SQL · PYTHON · PLOTLY

Financial Market Risk Analytics

VaR / CVaRMonte Carlo 10KSharpe · Sortino

End-to-end quant risk system on S&P 500 sector ETFs — Historical & Cornish-Fisher VaR, CVaR/Expected Shortfall, 10K GBM Monte Carlo simulation, rolling Sharpe/Sortino/Calmar, macro regime analysis (VIX, yield curve). Advanced PostgreSQL window functions + Python modelling.

PostgreSQLPythonPandasSciPyPlotlyyfinance

LIVE DASHBOARD

NEW SQL · ML · PLOTLY

E-Commerce Customer Analytics

RFM SegmentsROC-AUC 0.84100K+ Orders

Full analytics pipeline on Olist Brazilian E-Commerce (100K+ orders, 8 tables) — RFM segmentation, cohort retention, CLV estimation, churn prediction (Logistic Regression, AUC 0.84), K-Means clustering. Raw SQL to ML to interactive dashboard. Champions drive 34% of GMV.

PostgreSQLPythonScikit-learnPlotlyK-Means

LIVE DASHBOARD

CV / PYTORCH

Tesla-Style YOLOv5 Object Detector

62.4ms InferenceCustom Classes

YOLOv5 pipeline with custom class filtering for autonomous vehicle perception. Latency benchmarking and optimization targeting automotive sensor performance requirements.

YOLOv5PyTorchComputer VisionPython

Who I Am.

Where I've Worked.

Cipher & Entropy.

What I've Built.