yacine-ai-tech/yacine-ai-tech

What I Do

I work across three registers, and I try not to blur them.

Background

Open-Source Ecosystem — 6 tools, 2026, dual-licensed (AGPL-3.0 / commercial)

Each tool ships with a BENCHMARK.md and a RESEARCH.md in its repository — the evaluation protocol and the raw numbers, favorable and unfavorable both. Summaries below; the repos are the source of truth.

Click any Measured results panel to expand it.

IntelAI — Persona-Aware Enterprise Analytics & RAG Copilot

Nine C-suite personas (CEO, CFO, CTO, COO, CHRO, ESG, Risk, Analyst, General), each scoped to its own data domain by architecture, not prompt instruction. 146 curated KPIs across 7 domains, 78-month history (2020-01 to 2026-06). Hybrid retrieval (BGE-M3 dense 1024-dim + BM25 sparse + RRF fusion, with optional BAAI/bge-reranker-v2-m3 cross-encoder reranking applied progressively), plus GraphRAG-lite multi-hop entity traversal and prompt caching with dynamic recency cutoff (June 2026). The copilot runs rule-based live web search in parallel with retrieval for questions about external companies and current events, lists knowledge-base and web sources separately, and streams answers with visible pipeline stages; retrieval is bilingual (French/English) with a soft language preference.

DocIntel — Vision-First Document Intelligence

Extracts structured data from PDFs, images, and scanned files across three production routes: Route A (Frontier Multimodal Vision: Claude Sonnet 4.6 / Gemini 3.8 Flash Native Vision), Route B (Self-Hosted Vision Model via Ollama: Qwen 2.5-VL 7B with zero cloud API cost), and Route C (layout-aware OCR, Surya or Tesseract as configured, then LLM structured cleanup). Deterministic currency normalization across 45+ currencies (including FCFA/XOF under UEMOA VAT convention) and ISO dates, with SHA-256 document deduplication and sub-second cached retrieval.

RAGeval — Self-Hosted LLMOps Observability for RAG

from rageval import track

@track(project="my_rag_app")
async def answer(question): ...

Zero-credit symbolic evaluation combined with a heterogeneous 4-judge consensus panel (Groq gpt-oss-120b, Gemini 3.5 Flash, Claude Haiku, GPT-5-mini), rate-limiting, and persistent SQLite/PostgreSQL disk caching.

AgentKit — Governed MCP Tool Server

Gives Claude Desktop, Cursor, or any LangGraph agent governed access to live business data through the Model Context Protocol — typed tool effects (read/write/destructive), capability policies, and deterministic audit logging.

StreamPulse — Real-Time Business Data Pipeline

Multi-source ingestion (JSON, CSV, email, HMAC-verified webhooks, Google Sheets) with first-class n8n automation integration (custom node + 5 workflows) and a 3-tier hybrid classification cascade (regex/heuristics → semantic embeddings → LLM reasoning escalation).

VoiceFlow — Speech to Structured Business Intelligence

Routes recorded audio to per-analysis-type LLMs with a multi-provider transcription fallback chain (Groq Whisper, Deepgram, AssemblyAI, local WhisperX) and bidirectional real-time audio streaming via Gemini Live (gemini-2.5-flash-native-audio-preview-09-2025 with google-genai v1beta SDK, 24kHz → 16kHz downsampling, gated frame pipeline).

Results at a Glance

Client Work (scoped to what’s publicly shareable)

Research Direction

Live GitHub Activity