Overview documentation for my projects. Source code is private (not published); most projects are proprietary. These pages describe what each system does and how it is built, without the implementation. Code review available on request.
Contact: caciolo.edoardo@gmail.com, github.com/Ed0C97, linkedin.com/in/edoardo-caciolo
Projects come in two kinds, kept separate on purpose:
- Projects with code samples carry a folder with a short README and a few IP-safe excerpts, adapted and trimmed so each file reads standalone. The excerpts show the engineering (patterns, algorithms, resilience), never the proprietary core.
- Additional projects are described at the design and capability level, with no code sample published. A sample or review can be arranged on request.
Images referenced by any page live in projects/images/<project-slug>/, never beside the code samples, so a code folder stays code only.
The published excerpts are chosen to show concrete engineering, each tied to the project that evidences it:
- Information retrieval and ranking: hybrid dense-plus-sparse retrieval, Reciprocal Rank Fusion, and a cost-aware reranking funnel that degrades gracefully when a model provider is slow or down (ARBOR).
- Control and estimation math: Unscented Kalman Filter state estimation and LPV discretization for a vehicle control stack (thesis); numerically stable Erlang-B and Erlang-C queueing for capacity planning (ClusterMind).
- Native Apple platforms: real-time camera-quality signal processing and augmented-reality lifecycle handling in Swift (Tay); hand-laid PDF and CSV export (Tappo).
- Resilience and operability: circuit breakers on a monotonic clock, multi-backend secret resolution, and graceful-degradation fallbacks (Sentinel, fast-inference).
- Security-conscious design: SSRF and DNS-rebinding defense on a local gateway, injection-safe SQL over a recursive CTE, and fail-closed field encryption (BOT Garage, Alatar).
- Threat detection and intelligence: channel-isolated SIEM alerting in ArcSight CEF that never lets one dead sink block the others (NETWATCH); cross-source entity resolution into a knowledge graph and read-only-by-construction query safety (MINERVA).
- AI evaluation: dependency-free IR metrics shared by an eval harness and its metrics exporters so reported and dashboarded numbers cannot drift, and a judges-off-versus-judges-on benchmark that gates a verification layer on accuracy gain, false-positive reduction, and latency overhead together (Sentinel).
- Trustworthy data capture: confidence-gated OCR field extraction that persists a field only above a confidence threshold and routes weaker matches to human review (BOT Garage).
- ARBOR: A multi-tenant AI platform that discovers, enriches, and reasons over real-world entities, combining a knowledge graph, vector search, and an event-sourced data store. The published excerpt is its search and ranking engine: hybrid retrieval with RRF fusion and a cost-aware reranking funnel that degrades gracefully when a model provider is slow or down.
- Sentinel: A multi-tenant document intelligence platform for credit and risk teams that extracts, verifies, scores, and attests risk in due-diligence documentation. Every finding passes a grounding and coverage gate before acceptance, and reproducibility is enforced end to end: batch-invariant deterministic serving under multi-tenant concurrency, and extraction measured against golden sets in CI.
- Alatar: A career-intelligence platform that aligns a candidate's CV with a job description and returns a scored, evidence-backed match that both recruiters and candidates can audit.
- Tay: A native Apple-platform app that recognizes physical artworks and monuments through the camera and anchors real-time augmented-reality experiences on top of them.
- fast-inference: A self-hosted, OpenAI-compatible inference server for embedding, reranking, and generation models: INT8 static quantization, hand-written Triton GPU kernels, asynchronous dynamic batching, and a benchmark harness that discovers what the current machine can run and attaches the hardware to every measurement. Ships a self-contained operator dashboard.
- ClusterMind: A command-line tool that turns a short description of a web, SaaS, or self-hosted LLM service into a deterministic, auditable infrastructure capacity and cloud-cost plan.
- Hybrid Multi-Layer Control for Robust Autonomous Driving: A bachelor's thesis and simulation framework for a layered learning-plus-model-based control stack that keeps a Formula Student driverless race car safe and accurate when tire grip and vehicle parameters are uncertain.
| Project | What it is |
|---|---|
| Anacleto | A FastAPI platform that analyzes Instagram images with vision models, optimizes feed order via MCTS, Beam Search, or Greedy strategies, schedules posts, and publishes through the Meta Graph API. |
| Alatar | A career-intelligence platform that aligns a candidate's CV with a job description and returns a scored, evidence-backed match that both recruiters and candidates can audit. |
| ARBOR | A multi-tenant AI platform that discovers, enriches, and reasons over real-world entities, combining a knowledge graph, vector search, and an event-sourced data store. |
| BOT Garage | A local-first workshop management application that tracks vehicle maintenance, costs, and documents, with OCR import of data straight from registration and invoice PDFs. |
| ClusterMind | A command-line tool that turns a short description of a web, SaaS, or self-hosted LLM service into a deterministic, auditable infrastructure capacity and cloud-cost plan. |
| fast-inference | A self-hosted, OpenAI-compatible inference server for embedding, reranking, and generation models, built to keep retrieval-augmented generation (RAG) workloads fast and fully on-premise. Ships an operator dashboard and a benchmark harness that reports the machine behind every number. |
| MINERVA | An API-first OSINT engine that aggregates public cyber-threat data into a knowledge graph and answers natural-language questions about the threat landscape. |
| NETWATCH | A Linux endpoint security monitor that observes process, network, and file activity at the kernel level with eBPF, scores it against rules and machine-learning models, and dispatches alerts to a dashboard and SIEM systems. |
| Porfirio Magazine | A full-stack digital magazine platform that unifies bilingual publishing, payments, engagement analytics, and automated social-media content generation in a single product. |
| Sentinel | A multi-tenant document intelligence platform for credit and risk teams that extracts, verifies, scores, and attests risk in due-diligence documentation, producing grounded, reproducible findings instead of free-form summaries. |
| Tappo | A native iOS app for tracking counts, habits, and goal progress across multiple projects, with Home Screen widgets and Lock Screen Live Activities that stay in sync with the app. |
| Tay | A native Apple-platform app that recognizes physical artworks and monuments through the camera and anchors real-time augmented-reality experiences on top of them. |
| Hybrid Multi-Layer Control for Robust Autonomous Driving | A bachelor's thesis and simulation framework for a layered learning-plus-model-based control stack that keeps a Formula Student driverless race car safe and accurate when tire grip and vehicle parameters are uncertain. |
- Academic coursework: graduate-level interactive-graphics and machine-learning assignments (Sapienza University of Rome), implemented from scratch.
© 2026 Edoardo Caciolo, all rights reserved. Source code is private and available for review on request.