Our Projects
Full-stack AI infrastructure — from cloud-based content automation to on-device compliance and threat detection running on Apple Silicon.
SEO AI Agent
A full-stack AI platform that automates SEO content production — from a cloud-based orchestration engine to a modern management interface. The backend generates sitemaps, page content, metadata, schema markup, and redirect tables using orchestrated AI agents, while the frontend gives teams real-time visibility, editing, and approval control over every deliverable.
AI Agent Orchestration
CrewAI flows coordinate strategy and implementation agents in an L2/L3 task pattern — high-level decisions flow to mechanical executors, reducing hallucination and ensuring consistent output quality.
Real-Time AI Chat
Server-Sent Events streaming delivers live agent progress in the management UI — tool calls, thinking steps, and final responses appear in real time with heartbeat keepalive.
SEO Artifact Generation
Automated production of sitemaps, redirect tables, metadata, schema markup, llms.txt, and page content — configured per client, generated per page, ready for review.
Review & Approval Workflow
Browse, edit, and approve deliverables with version history. Shareable review tokens let external stakeholders view and approve artifacts without an account.
Intelligent Tool Suite
Sitemap discovery, page retrieval with HTML-to-Markdown conversion, full website crawling, marketing profile extraction, business data scraping, and review summarization.
Observability & Budget Control
Langfuse + OpenLit tracing, per-job token tracking, dollar-based budget enforcement, and Cloud Run job orchestration with automatic dispatch.
AgentCheck
An on-device AI-powered compliance auditing app for iOS/macOS that analyzes audio recordings of customer service calls and evaluates whether agents meet regulatory criteria — identity disclosure, fee transparency, right of rescission, and privacy agreements. All speech recognition and text analysis runs locally on Apple Silicon via MLX, with zero cloud dependency.
On-Device Compliance Analysis
Upload a call recording, transcribe with on-device ASR, then run each compliance criterion through a local LLM. Per-criterion pass/fail with extracted evidence from the transcript.
Real-Time Live Streaming
Stream microphone audio in real-time, transcribe on-the-fly with streaming ASR, and automatically evaluate compliance criteria as speech accumulates — animated checklist progress included.
Configurable Domain Templates
Pre-built Financial Consultant and Insurance Agent templates with 6 criteria each. Add, edit, delete, or reset criteria per domain — persisted via AppStorage as JSON.
On-Device Noise Cancellation
Three modes for live audio: Off, DeepFilterNet V3 (MLX streaming, no API key), or Picovoice Koala (real-time SDK). Anti-aliased vDSP decimation prevents sibilant artifacts.
Robust LLM Response Parsing
Multi-layer JSON parser strips markdown fences, handles flexible field names and _met as bool/string/int, with regex fallback — ensuring reliability across model families and quantization levels.
Resource Monitoring
Real-time RAM and CPU tracking via Mach API calls (phys_footprint and thread_info), surfacing memory pressure during model loading and inference to prevent iOS out-of-memory terminations.
NoiseCancellation
An iOS application for comparing two on-device noise cancellation engines side by side. Record audio or import a file, process it through Picovoice Koala and DeepFilterNet-MLX simultaneously, then playback all three versions (original, Koala, DeepFilterNet) for direct A/B/C comparison — all processing runs locally on Apple Silicon.
Side-by-Side Comparison
Record once from the microphone, then listen to three outputs in parallel: raw original, Koala-processed, and DeepFilterNet-processed. Auto-stop cross-player controls keep playback intuitive.
Real-Time Koala Processing
Picovoice's Koala SDK processes each audio frame in real-time during recording — frame-by-frame noise suppression at 16 kHz, 16-bit PCM, mono. No post-processing wait.
On-Device DeepFilterNet Inference
DeepFilterNet3 model (2.1M parameters) runs via Apple MLX on Apple Silicon. Supports both streaming mode (low latency, 10ms chunks) and batch mode (higher quality, full-file processing).
Audio File Import
Import WAV, MP3, or M4A files via system picker for batch processing through the DeepFilterNet engine — compare quality without live recording.
Configurable Processing
Adjustable attenuation limit slider (0–100 dB) for DeepFilterNet aggressiveness. Streaming mode toggle for real-time vs. batch. Settings persist across app launches.
Peak-Normalized WAV Output
All processed audio is peak-normalized to 0.9 amplitude before saving to WAV, ensuring consistent playback levels across original, Koala, and DeepFilterNet outputs for fair comparison.
ScamDetector
An iOS app that detects potential scam conversations from audio recordings. Upload a phone call recording, get an on-device ASR transcription, then an on-device LLM analysis that identifies common scam patterns and returns a risk score (1–5) with a detailed explanation — all processing stays on the device, nothing leaves the phone.
Audio Upload & Transcription
Import audio files (m4a, mp3, wav, aac, flac) via native iOS picker. Long audio is chunked into 12-second segments with 1s overlap and merged with overlap-aware deduplication to stay within memory limits.
LLM-Powered Scam Analysis
On-device LLM analyzes transcripts for scam patterns — government impersonation, urgency tactics, gift card requests, elderly targeting — and outputs a risk score (1–5) with detailed explanation.
Configurable Model Selection
Choose from multiple ASR models (Polyglot Lion 0.6B/1.7B, Qwen3 ASR 0.6B) and LLM models (Qwen3 0.6B–4B, Qwen3.5 0.8B–4B) at various quantizations. Downloaded once, cached on-device.
Customizable System Prompt
Edit the LLM system prompt in-app with a [TRANSCRIPT] placeholder. Full control over where the transcript is inserted and how the analysis is framed. Reset to default anytime.
Live Processing Pipeline
Step-by-step pipeline UI with expandable cards that auto-expand as each stage runs — real-time transcription progress, streaming LLM output, and color-coded risk score display (green/orange/red).
iOS Memory Optimization
Chunked audio processing (12s segments with 1s overlap), autoreleasepool scoping, and MLX GPU cache clearing between inference runs keep memory usage within iOS limits on device.