Most AI systems fail somewhere between a working demo and a working system. Retrieval that drifts, agents that don't recover, evaluation that catches nothing. I work in that gap.
Backend systems, retrieval, agents, evaluation. I build reliable AI systems for teams that have outgrown the prototype phase: RAG with measurable retrieval quality, agent workflows with state and approval gates, FastAPI backends purpose-built for LLM products, and the evaluation layer most teams skip between prototype and production.
Available for contract work. ignaziodesantis.com · eleventh.dev · LinkedIn · Email
Stack: Python · FastAPI · LangGraph · pgvector · Anthropic Claude · React · TypeScript · Next.js · Tauri · Rust
| System | What it is |
|---|---|
| Orion | Single-node, operator-supervised AI freelance agent powered by the Anthropic Claude API. End-to-end execution pipeline (Scout → Qualify → Propose → Execute → Deliver → Follow-up) across three tracks: enterprise, engineering, webdev. Closed-loop intelligence layer: prompt-drift detection, active-learning queue, Monte Carlo deal predictions, Bayesian pricing calibration, per-LLM-call cost attribution. Python 3.12 · FastAPI · SQLite · Anthropic SDK · React · Stripe. Private repo. |
Four engineering capabilities. Each row links to its source repo.
| # | Capability | System |
|---|---|---|
| 01 | Retrieval Contracts for RAG Systems — Hybrid retrieval (pgvector + BM25) with reranking, instrumented end-to-end. Built for the gap between a demo that retrieves and a system that survives the next 10,000 queries. | NexusRAG |
| 02 | LLM Agent Infrastructure — Multi-step agent workflows on LangGraph with persistent state, tool orchestration, approval gates, retry logic, and deterministic evaluation. | agent-runbook-orchestrator |
| 03 | FastAPI Backends for AI Products — Async FastAPI backends purpose-built for LLM-powered apps. pgvector, Alembic migrations, health checks, operational readiness on day one. | langgraph-fastapi-starter |
| 04 | AI Reliability Engineering — Deterministic evaluation frameworks, structured output validation, cost and latency observability, regression detection. The layer most teams skip between prototype and production. | evalops-workbench |
Plus: SentinelID — passkey-style face authentication system. On-device CV for identity verification, liveness detection, and anti-spoofing. Desktop app (Tauri/Rust) + web dashboard (Next.js + FastAPI).
| Repo | Description |
|---|---|
| evalops-workbench | Local-first evaluation harness for prompts, tools, and agents with regression tracking |
| repo-rag-debugger | Source-aware debugging assistant that indexes codebases and docs to propose grounded fixes |
| spec-to-agent-scaffold | Generates typed Python agent service skeletons from structured product specs |
| Repo | Description |
|---|---|
| data-quality-watchtower | Schema drift detection, anomaly monitoring, and dataset validation before pipelines break |
| notebook-pipeline-converter | Converts exploratory notebooks into repeatable, testable batch pipelines |
| api-sync-pipeline | Incremental REST API to SQLite sync engine — cursor pagination, retry/backoff, YAML config |
| marketing-attribution | Multi-touch attribution analysis: last-click vs linear vs time-decay across 17k touchpoints |
| churn-analysis | SaaS customer churn cohort analysis — onboarding signals, channel quality, retention heatmaps |
| Repo | Description |
|---|---|
| revenue-signal-copilot | Lead intelligence and prioritization using public signals and internal notes |
| customer-ops-briefing-bot | Weekly account intelligence agent from CRM activity, meeting notes, and support history |
| lead-aggregator | Multi-feed RSS lead aggregator with keyword scoring, deduplication, and CSV/SQLite export |
| price-monitor | Multi-site e-commerce price monitoring with threshold alerts and CSV reporting |