The chat is for information only and does not offer professional advice.
Hello world, my name is Eric Allione. I take AI systems from empty repo to production, solo and fast.
In the last three months I designed and shipped roughly ten production systems from first principles, serverless and framework-free, in Rust, Python, and vanilla JavaScript: a real-time options and volatility quant platform, self-updating RAG knowledge bases, and a publication-grade verified research archive. The range comes from a method, not a stack: first-principles reasoning trained in philosophy, amplified by parallel agentic workflows that turn any bounded domain into a solvable problem.
Production systems, built from scratch.
A sample of what I have shipped solo in the last three months. Every one is hand-coded, serverless, and template-free, with the math and the infrastructure built rather than imported. Live links where the work is public.
A real-time S&P 500 options and volatility dashboard with nine in-browser quant labs running SVI, Heston, gamma-exposure, and volatility-risk-premium models over years of reconstructed market history, fed by a self-hosted end-of-day pipeline and a RAG-grounded assistant.
An educational philosophy reference where 156 thinkers each get a sourced, idea-first study room and an AI guide grounded in public-domain primary texts, linked by an interactive similarity graph and a vintage geo-coded world map, kept current by a nightly ingestion pipeline.
An AI news and learning platform on a fully automated pipeline that pulls from seven sources, classifies and summarizes each item with AI, and embeds it into live semantic search, beneath a hand-curated glossary of illustrated concept pages with automatic cross-linking.
A Rust learning platform written end-to-end in Rust, from server and browser to database and deployment, with an in-browser code runner and an AI tutor grounded in its own curriculum. The build verifies itself before it ships.
A companion browser extension that surfaces the aigamma.com market-regime read directly in the toolbar: a live status icon backed by minimal-permission Manifest V3 service workers, CORS-scoped fetches, and market-hours-gated caching against custom Netlify functions.
A public site for testing rules-based stock-sector strategies against real historical data, with six strategies, a written curriculum and quiz, and SelectBot, a from-scratch streaming teaching chatbot that remembers each visitor's conversation between visits.
From the model to the metal.
I work top to bottom: the AI architecture that grounds models in verifiable data, the systems engineering that makes it fast and observable, and the quantitative analytics underneath. The through-line is correctness enforced by oracles, not assurances.
AI Architecture
- Custom RAG pipelines: vector retrieval with keyword and tsvector fallback, rate limiting, and self-refreshing corpora via cross-repo CI
- Cross-provider multi-model verification: a drafter plus an independent skeptic, with a human-in-the-loop gate that auto-holds unsupported claims
- MCP servers and Claude Desktop integration exposing private corpora to external AI tools
- Streaming SSE assistants and AI tutors grounded in primary-source and curriculum content
- Vector databases: Qdrant, Pinecone, and Supabase pgvector with HNSW and embedding migration
- Prompt-injection defense and AI data-provider guardrails
Systems Engineering
- Self-hosted scheduled data pipelines with retries, exponential backoff, and idempotent corrections
- Build-gating test suites, from 200 to nearly 1,000 checks, failing on broken links or out-of-sync docs
- Performance work: two-thirds payload cuts, 3.4s to 0.4s page loads, a 19-step fetch rewritten to a sub-100ms query
- Distributed observability on OpenTelemetry, Prometheus, and Grafana, with PII stripped before data leaves the servers
- Rust end-to-end and Rust-to-WASM compute cores, Tauri desktop apps, and cross-store browser extensions
- Cross-repo CI/CD on GitHub Actions, per-IP rate limiting, and documented cost defenses
Backend & Analytics
- Quantitative modeling: SVI surface fitting, Heston and Black-Scholes pricing, gamma exposure, and volatility risk premium
- Backtesting engines with rules-based strategies and anti-bias statistics (Bonferroni, Holm, Benjamini-Hochberg)
- Postgres and Supabase data layers with timestamped audit trails and automated verification against authoritative references
- Row-level security enforcing access control at the database under the caller's identity, not in app logic
- Inter-coder reliability statistics, Krippendorff's alpha and Cohen's kappa, verified against textbook examples
- Word-level transcript reconciliation and large-scale historical ingestion
The numbers behind the work.
A spread of the strongest figures, each one verifiable in a live product, a commit history, or a documented record. Recent build metrics sit alongside the quantitative-finance and defense work that came before.
Brought in to fix what mattered.
A selection of client and contract work across heritage archives, academic research, regulated healthcare, travel, and finance. The most recent engagements are access-gated to protect interview confidentiality or pending institutional approval, so they are described here without names or links.
Accuracy is enforced by a cross-provider dual-model pipeline: one model drafts, a second skeptical model from a different provider independently checks every factual claim, and anything unsupported is auto-held at a publication gate for human review. Transcripts were reconciled word by word against canonical source recordings, with deep-linkable cited search, audio-seek playback, and MCP integration layered on top.
Hybrid keyword and semantic search with reranking runs over the interview corpus with row-level security, so confidentiality is enforced at the data layer rather than in app logic. It grew into a fleet of instruments, including a constellation view of thousands of passages and an inter-coder reliability module computing Krippendorff's alpha and Cohen's kappa, all built from precomputed database relationships.
Audited the vendor stack and tracking infrastructure, exposed a click-hijacking network affecting the large majority of promised ad delivery, stripped redundant tag-manager containers and conflicting third-party integrations, and rebuilt campaign configuration to outperform default automated offerings. Also devised an automated currency-hedging approach that let leadership lock long-term forward bookings.
Stood up the compliant environment, then delivered a library of discrete operational workflows across patient intake, scheduling, programmatic content generation, and retention tracking, pairing local inference nodes for confidential data with secure API boundaries for secondary processing.
Provisioned launch logistics including overnight distributed cash sweeps, compiled a parameterized alpha risk-management playbook, engineered multi-GPU tensor-parallel inference for bounded financial tasks, and instituted repository controls, CI/CD checks, and review harnesses to turn loose research into repeatable R&D infrastructure.
Multiple domains. One methodology.
I started in the defense sector as acting lead infrastructure architect for a defense agency spanning 22 countries, holding a Top Secret clearance. Hardened systems and documentation discipline became the operating layer for everything that followed. I then spent five years in full-time independent study of quantitative finance, working through hundreds of primary texts on derivatives theory and volatility modeling while trading derivatives professionally.
That self-directed education led to a principal role at a derivatives analytics firm, where I became the sole authority on model correctness and ghostwrote 585 institutional market-close newsletters and a 700-page knowledge base, developing original theory on gamma exposure and volatility-surface dynamics in real time under daily deadlines, without generative AI. I answered over 10,000 technical questions from professional traders and built the volatility tools that taught them.
Underneath all of it is philosophy. I have studied the Western tradition seriously since I was fifteen, earned a B.A. in Philosophy with a Certificate in German to read primary texts in the original, and still work through major works actively. That training is not ornamental: systematic philosophy is how I decompose a problem to first principles, surface unstated assumptions, and build frameworks from the ground up, which is why I can enter an unfamiliar domain and operate at a high level quickly.
The practical expression of that is how I use AI. I run parallel inference across an ensemble of engines, surface and cross-pollinate minority opinions, and relay one model's criticism to another, governing a reasoning process rather than outsourcing it. The philosophy supplies the reasoning architecture; the AI ensemble supplies the execution speed. Together they are what let one person design, build, and harden production systems across finance, AI, regulated healthcare, and research at the pace shown above.
B.A. Philosophy (Certificate in German), UH Manoa
CASP+, CNDA, C|HFI (Inactive)
Capstone Certificate of Excellence
Golden Key International Honour Society
King's College London
Let's build something.
Open to senior and staff-level AI engineering and architecture roles, plus fractional and contract engagements and hard problems that need a system taken from idea to production. Remote, worldwide. The fastest path is a direct conversation.
Schedule a Meeting Download Résumé