Olayinka David Vaughan

MARL economy · live shared mainframe

Multi-Agent Reinforcement Learning desktop environment for decentralized economic simulation. PPO-trained agents discover pricing, wage-setting, and consumption strategies inside a glassmorphic OS-style dashboard.

Cluster: quant. Stack: Python, PettingZoo, PPO, FastAPI, WebSocket.

• Live: https://econ-os.vercel.app
• Repository: https://github.com/Builder106/EconOS

~18M orders/sec · p99 < 1µs

High-performance limit-order-book matching engine in OCaml 5. Allocation-free per-submit hot path, ~18M orders/sec, p99 latency under 1μs. Dream HTTP + SSE backend with a Bloomberg-terminal-style browser dashboard.

Cluster: quant. Stack: OCaml, Dream, Docker, Tailwind, Oracle Cloud.

• Live: https://ocaml-lob.vercel.app
• Repository: https://github.com/Builder106/ocaml_limit

Neural net in ~2k LoC C99 · zero deps

A neural network built from scratch in C99 — no TensorFlow, no PyTorch, no dependencies. Trains on market data, runs a DQN trading agent that outperforms buy-and-hold, ships as a WebAssembly demo.

Cluster: quant. Stack: C99, WebAssembly, Emscripten.

• Live: https://qforge-neural.vercel.app
• Repository: https://github.com/Builder106/qforge

123 tests passing · NGO ↔ volunteer marketplace

A micro-volunteering marketplace pairing NGOs with volunteers for 5–30 minute skill-building tasks. Volunteers browse the feed, claim missions, submit proof, and earn badges. NGOs post tasks and review submissions.

Cluster: SWE. Stack: SvelteKit, Appwrite, Bun, TypeScript, Tailwind.

• Live: https://trymicromatch.vercel.app
• Repository: https://github.com/Builder106/MicroMatch

Live applicant flow · Nigeria's STEM scholars

Web platform for STAIJA's StepUp Scholars and Dynamerge programs — applicant tracking, mentorship workflow, and public content for Nigeria's STEM students. Application management, role-aware routing, and Mailgun-backed comms.

Cluster: SWE. Stack: Vue 3, TypeScript, Firebase, Tailwind, Vite.

• Live: https://staija.org
• Repository: https://github.com/Builder106/STAIJA

Focus → plants → leaderboard

A study tracker that turns focus sessions into a growing garden. Run a focus timer, log sessions toward a daily goal, watch plants grow over time, and compare streaks on a public leaderboard.

Cluster: SWE. Stack: Deno 2, React, TypeScript, Vite, Supabase.

• Live: https://getstudysprint.vercel.app
• Repository: https://github.com/Builder106/StudySprint

16,203 trades · +2.58% alpha (p < 0.05)

End-to-end Python pipeline auditing 16,203 disclosed Congressional stock trades from 2020–2024. Scrapes Senate and House Periodic Transaction Reports with Playwright + pdfplumber, computes Jensen's alpha against the S&P 500, and ships a Vercel findings page. Semester project for Wesleyan's QAC 420 — Data for Good.

Cluster: analyst. Stack: Python, Playwright, pdfplumber, pandas, scipy, Jupyter.

• Live: https://capitolalpha.vercel.app
• Repository: https://github.com/Builder106/CapitolAlpha

3× ED-visit odds · n = 58,639 patients

ASA DataFest 2026 submission for Stormont Vail Health. R + DuckDB pipeline on a 7.6M-row EHR sample joined to a social-determinants questionnaire — patients reporting a transportation barrier show ~3× crude odds of ED visits and inpatient admits, independent of age. Wesleyan team 13.

Cluster: analyst. Stack: R, DuckDB, data.table, ggplot2, Flourish.

• Live: https://datafest-2026.vercel.app/
• Repository: https://github.com/Builder106/datafest-2026

Phoronix Test Suite · monthly CI captures

A benchmarking dataset and Rails 8 dashboard comparing Ubuntu, Fedora, and Debian under identical virtual hardware. Phoronix runs are captured monthly by GitHub Actions; R parsers and the dashboard consume the same composite XML, so the static analysis and the live UI never drift.

Cluster: analyst. Stack: Ruby on Rails, R, Phoronix Test Suite, GitHub Actions, Docker.

• Repository: https://github.com/Builder106/LinuxBenchHub

Source-rebuild gatekeeper · blocks supply-chain tampering

A pre-install gatekeeper that answers “is this binary actually a build of the source it claims?” Fetches a package, verifies its SLSA attestation through Sigstore + Rekor, rebuilds it from the attested source commit in an isolated Docker container, and blocks the install if the rebuilt file tree diverges from the registry artifact. Catches the event-stream / ua-parser-js / xz-utils class of attacks.

Cluster: cybersec. Stack: Rust, Sigstore, Rekor, Docker, Axum.

• Live: https://clear-hash.vercel.app
• Repository: https://github.com/Builder106/ClearHash

MCP firewall · p50 ≤ 200µs · 50k req/s

A JSON-RPC firewall that sits between an LLM agent and its Model Context Protocol servers. Every tools/call envelope is parsed, evaluated against a YAML policy bundle, and either forwarded or blocked with a synthetic error before a malicious payload reaches the host — defending against tool poisoning, argument injection, capability creep, and secret exfiltration. The policy engine compiles to WebAssembly for an in-browser playground.

Cluster: cybersec. Stack: Go, WebAssembly, Next.js, JSON-RPC, MCP.

• Live: https://halberd-keep.vercel.app
• Repository: https://github.com/Builder106/Halberd

188-byte Yul executor · 99.89% arb prediction

A bare-metal MEV arbitrage engine scoped to cross-DEX back-running. A TypeScript mempool scanner spots swaps about to land on Uniswap-V2-shaped pools, solves the optimal back-run input against post-victim reserves, and — if profit beats gas — packs calldata for a 188-byte Yul executor funded by an Aave V3 flashloan. No inventory; predatory sandwich and JIT strategies are explicitly out of scope.

Cluster: cybersec. Stack: Solidity, Yul, TypeScript, Foundry, Bun.

• Live: https://quarry-mev.vercel.app
• Repository: https://github.com/Builder106/Quarry

0 B egress on-device · privacy measured per run

An interactive workbench for clinical-document extraction that never phones home. Load a synthetic superbill, pick an extractor — a deterministic rules parser, a local on-device model (qwen2.5:3b via Ollama), or a hosted cloud model (Groq) — and watch a structured record fill in field-by-field against ground truth while a live gauge measures whether the document's bytes stayed on-device (0 B) or crossed to the cloud. Privacy isn't asserted, it's measured per run over 50 held-out synthetic superbills.

Cluster: AI/ML. Stack: Next.js 16, AI SDK 5, Ollama, Groq, TypeScript.

• Live: https://enclave-iota.vercel.app
• Repository: https://github.com/Builder106/Enclave

99% invoice OCR · $0.0003/invoice · 15.4× faster

A Gemini 3.1 Flash Lite + MCP executive co-pilot for small-business operations. Four back-office workflows — AP-invoice OCR, creator-payout reconciliation, Tier-1 customer-service responses, and cross-company KPI Q&A — run end-to-end with measured cost and accuracy per task. AP-invoice OCR hits 99% parse / 91.9% field accuracy at $0.0003 per invoice (15.4× faster than a 6-min manual baseline); the payout reconciler exposes where an LLM reasons badly about multi-step arithmetic, so a deterministic re-computer disposes what the model proposes.

Cluster: AI/ML. Stack: React 19, Gemini 3.1 Flash Lite, MCP, libSQL, Node.

• Live: https://helm-bridge.vercel.app
• Repository: https://github.com/Builder106/Helm

RAG over wiki + Discord + market data · writes Trader classes

A retrieval-augmented assistant for the IMC Prosperity trading competition. Ensemble retrieval over three weighted vector stores — competition wiki, community Discord exports, and historical market data — grounds answers about products, position limits, and strategy, and generates complete, ready-to-run Trader classes. Groq-backed generation (llama-3.3-70b-versatile) keeps inference fast inside a Streamlit chat UI that cites its sources per answer.

Cluster: AI/ML. Stack: Python, Streamlit, Groq, RAG, vector search.

• Live: https://tradetell.streamlit.app
• Repository: https://github.com/Builder106/IMC_Prosperity