01
Model lifecycle
PAIN POINT
↓Provider deprecates model
RISK
Forced migration + revalidation
LOCAL-FIRST SOLUTION
Pin the model. Migrate on your timeline.
LOCAL-FIRST AI ARCHITECTURES
Building infrastructure
for product-grade local AI.
I explore how local-first architectures can give AI products more control over model lifecycle, costs, data, and runtime, without giving up frontier models where they still make sense.
Closed models should be an option, not a dependency.Local-first means having another execution path when control matters.
- Applications: ClosedRoom, RedactGuard, and Aura Finance.
- Infrastructure: Local LLM Server, Local ASR Server, and Android LLM Harness.
- Hardware: laptop, desktop, and Android devices, with no cloud required.
WHY LOCAL-FIRST
The problem isn't the cloud.It's depending on intelligence you don't control.
Closed APIs are powerful. The risk is making them your only execution path.
02
Cost dependency
PAIN POINT
↓Per-token API pricing at scale
RISK
Unpredictable recurring costs
LOCAL-FIRST SOLUTION
Shift suitable workloads to fixed capacity.
03
Data boundaries
PAIN POINT
↓Raw data sent to cloud APIs
RISK
Privacy risks & compliance friction
LOCAL-FIRST SOLUTION
Keep sensitive processing 100% on-device.
04
Exit optionality
PAIN POINT
↓Proprietary API lock-in
RISK
High switching cost on terms change
LOCAL-FIRST SOLUTION
Build alternative local execution paths.
Between Claude Opus 4.1's deprecation notice and API retirement in 2026.
Anthropic ↗Of surveyed Italian IT decision-makers say their organisation has an AI-provider exit strategy.
Red Hat ↗Qwen3.5-4B reports 79.1 MMLU-Pro with a 262k native context window.
Qwen ↗These signals don't make local AI the answer to every workload. They make architectural optionality increasingly worth evaluating.
LOCAL-FIRST ≠ LOCAL-ONLY
Use frontier models when you need capability,not because you have no other choice.
Local where sufficient
Cloud where valuable
Portable by design
CONTROL COMES WITH RESPONSIBILITY
Running a model locally is easy. Running it reliably is the real engineering problem.
Local-first architectures give you full control over models, data, and costs but they also move engineering responsibility back to the product team across four key domains:
01
Runtime lifecycle
Model management, hardware compatibility, dynamic switching & zero-downtime updates.
02
Observability & telemetry
Performance tracking, latency metrics, accuracy signals & memory profiling without privacy leakage.
03
Hardware adaptation
RAM/VRAM constraints, resident memory optimization & thermal management on desktop and mobile.
04
Developer ergonomics
Unified OpenAI-compatible APIs, reliable local IPC & intuitive user experiences.
THE INFRASTRUCTURE STACK (THREE PILLARS)
Three infrastructure pillars
Each project solves a different part of the local-AI product stack. Reusable infrastructure, not isolated experiments.
ACTIVE
Local LLM Server
PROBLEM
Local applications should not need to couple product logic to inference backend complexities, model formats or runtime processes.
SOLUTION
High-performance local inference server with model lifecycle, dynamic routing and telemetry.
Explore the pillar
Provides an OpenAI-compatible API across local backends such as GGUF and Apple Silicon MLX, managing resident runtimes, configuration and telemetry.
- OpenAI-compatible local API
- GGUF and Apple Silicon (MLX) backends
- Runtime model switching & multi-residency
ACTIVE
Local ASR Server
PROBLEM
Meeting audio is highly sensitive, yet speech products routinely offload raw recordings to remote cloud infrastructure.
SOLUTION
Privacy-first speech recognition server optimized for sub-second latency and accuracy.
Explore the pillar
Handles microphone & system audio capture, local session storage, and Whisper transcription on Apple Silicon with zero cloud calls.
- Microphone & system-audio capture
- Local Whisper transcription with VAD
- OpenAI-compatible audio endpoint
IN PROGRESS
Android Local LLM Harness
PROBLEM
Running LLMs on Android demands rigorous resource management: model integrity, memory bounds, prompt cancellation and telemetry.
SOLUTION
On-device LLM runtime harness for Android apps with strict resource control.
Explore the pillar
Reusable runtime for embedding GGUF models in native Kotlin and Capacitor Android applications with lifecycle control and diagnostics.
- GGUF import & cryptographic integrity check
- Managed model and context lifecycle
- Real-time streaming & cancellation
ARCHITECTURE RELATIONSHIP
Local LLM ServerLocal ASR ServerAndroid LLM Harness
APIs / Protocols (gRPC, WebSocket, Local IPC)Reference Applications (ClosedRoom, RedactGuard, Aura Finance)
PROVING GROUNDS
Reference applications that validate the stack
ClosedRoom, RedactGuard, and Aura Finance demonstrate how privacy-first AI protects conversations, sensitive documents, and personal financial data in practice.
ClosedRoom →
Privacy-first meeting intelligence.
Local-first meeting capture, Whisper transcription and structured insights. Your conversations stay 100% on-device.
TranscriptionSummariesAction ItemsSearch
What it validates
- Local ASR & private speech recognition
- Local LLM analysis & structured insights
- Installable desktop app packaging & offline workflows
RedactGuard →
Local document anonymization & PII protection.
A local document-anonymisation workflow that detects PII via local LLMs, provides human-in-the-loop review, and selectively redacts sensitive data. Data minimisation built into the architecture.
Doc IntelligencePII DetectionHuman ReviewSelective RedactionSafe Export
What it validates
- Local document intelligence & PII detection
- Human review & selective AI redaction
- Data minimisation & safe document portability
Aura Finance →
Privacy-first personal finance.
A local-first personal finance application where transactions, budgets, reports and preferences live on-device first, with optional encrypted cloud backup.
Local DataReportsBudgetsEncrypted Backup
What it validates
- Privacy-first product design & local data model
- Mobile UX & local storage sovereignty
- Future proving ground for Android on-device AI integration
Reference applications validate the infrastructure and show what it can enable. They are the beginning, not the limit, of the ecosystem.
PROFESSIONAL TRACK RECORD
Research backed by real-world AI delivery
Local-first research and engineering built on years spent evaluating, designing, and scaling AI systems in complex enterprise environments.
Enterprise
AI Technical Leadership
Prototype → production
Experience leading AI initiatives in complex organisations, turning experiments into adopted, governed, and scaled systems.
80+
Articles Published
Research & insights
Long-form research, model evaluations, and practical LLM insights shared with the technical community through stAI tuned.
150+
AI Projects Reviewed
Technical mentoring
Hands-on feedback, architecture guidance, and mentoring for emerging AI engineers and technical teams.
Top Nova 111
Nova Talent
Recognition
Selected by Nova Talent among emerging Italian talent driving innovation and technology leadership.
KNOWLEDGE & PUBLIC RESEARCH
Building, documenting, and sharing in public
Experiments become code, measurements, technical notes and public discussion:Build → Measure → Document → Share.
ABOUT DANIELE
I work where AI infrastructure, product strategy, and clear engineering intersect.
I am Daniele Moltisanti, an AI technical leader and builder based in Milan. My work combines technical architecture, product thinking, and the ability to transform complex AI research into understandable decisions and reliable, intuitive software.
Alongside my independent work, I lead and design AI initiatives in complex organisations, translating technical possibilities into systems that can be adopted, governed, and scaled.
I also write, teach, and speak about AI architecture, strategy, and the path from experimentation to reliable products.
COLLABORATE
Exploring whether part of your AI stack
should move local?
I'm interested in real workloads: benchmarking a closed-vs-local task, testing models and devices, designing hybrid architectures, or exploring privacy-sensitive AI workflows.
danielemoltisanti@gmail.com Milan, Italy