I build production AI systems — the kind that get deployed, stay deployed, and solve a specific problem well.

Six standalone tools shipped in 2026. All open-source (AGPL-3.0). All live. Four on PyPI.

The core idea: different executives need different answers from the same data. A CFO asking about margins gets a Finance-scoped response. A CHRO asking about turnover gets a People-scoped one. The scope boundaries are enforced by architecture, not prompt engineering.

• 9 C-suite AI personas — CEO, CFO, CTO, COO, CHRO, ESG, Risk, Analyst, General — each with its own data scope, system prompt, and tool whitelist
• 169 curated KPIs across 7 domains (Finance, HR, IT, Ops, Logistics, ESG, Risk) — 36-month history, 7 benchmarking scenarios
• Hybrid retrieval — BGE-large-en-v1.5 dense + BM25 + RRF fusion + BGE Reranker v2-m3 cross-encoder; Cohere rerank as hosted backstop (resilience when inference Studio is unreachable)
• GraphRAG-lite — entity graph over KPI records for multi-hop queries (USE_GRAPH_RAG=true)
• WebSocket streaming with source citations; Qdrant in prod, Chroma in dev, pgvector available via env
• ML forecasting with Monte Carlo confidence bands; four-method anomaly detection; board-ready PDF export
• LiteLLM multi-provider router — Groq LLaMA 3.3 70B (default/speed), Claude Sonnet 4.6 (reasoning), Claude Haiku 4.5 (judge), Ollama (local)
• Tavily web search with cited sources (real-time data for the copilot)
• JWT + RBAC, bilingual EN/FR React frontend (Recharts), 7 benchmarking scenarios
• 74 endpoints, 157 tests, AGPL-3.0

pip install intelai  # v0.1.2

also: pip install omnismart-personas — v0.1.3 — standalone persona templates for LangChain RAG projects

FastAPI LangChain ChromaDB Qdrant pgvector PostgreSQL React Groq Claude LiteLLM Docker

Six MCP tools that give Claude Desktop, Cursor, or any LangGraph agent direct access to live business data.

• 6 MCP Tools: query_kpis, get_company_health, detect_kpi_anomalies, forecast_metric, list_available_metrics, get_executive_summary
• 6 MCP Resources (stable data URIs: kpi://Finance/latest, Growth, Operations, People, ESG, IT_Ops)
• 1 reusable Prompt — monthly_executive_briefing
• 3-agent LangGraph workflow — Planner (Claude Sonnet 4.6) → Analyst (Groq LLaMA 3.3) → Reporter (Claude Sonnet 4.6)
• Separate working demos for Claude Agent SDK, CrewAI, and DSPy
• 34 tests, AGPL-3.0

pip install agentkit-mcp  # v0.1.4
agentkit-mcp              # CLI entrypoint → starts the MCP server

FastMCP LangGraph Claude Sonnet 4.6 Groq PostgreSQL LiteLLM

Extracts structured data from PDFs and images using vision models — not just OCR.

• Route A: Claude Sonnet 4.6 Vision — complex layouts, handwriting, mixed languages
• Route B: Ollama local vision (.env default: llama3.2-vision; production-validated: qwen2.5-VL:7b on NVIDIA T4) — fully local, zero cost per page
• Route C: Tesseract + Claude Haiku LLM cleanup — lightweight fallback for clean scans
• Multi-currency & multi-locale — 45+ currencies (USD, EUR, GBP, JPY, INR, CNY, XOF/FCFA) normalized to ISO 4217 + float; dates to ISO 8601
• Multi-page map-reduce — handles 100+ page PDFs via concurrent chunk extraction and merge (MAX_PDF_PAGES default: 200)
• /classify-image — vision-first object classification (category + confidence + reasoning)
• 550-document benchmark (500 with field-level ground truth):

• 15 endpoints, 62 tests, AGPL-3.0

FastAPI LiteLLM Claude Vision Ollama Tesseract pdfplumber Docker

Record audio in the browser, get structured output — not just a transcript.

• 5 analysis types with per-type LLM routing:
  • meeting → Groq LLaMA 3.3 · sales_call → Claude Sonnet 4.6 · support_call → Claude Haiku 4.5 · interview → Claude Sonnet 4.6 · general → Groq LLaMA 3.3
• 4 transcription providers: WhisperX (local default) · Groq Whisper · Deepgram · AssemblyAI
• Diarization fallback chain: pyannote 3.x → NeMo → no-diarization
• Real-time voice agent via OpenAI Realtime API; when only GEMINI_API_KEY is set, routes to Gemini Multimodal Live automatically (translation layer in api.py:338)
• 13 endpoints, 38 tests, AGPL-3.0

FastAPI WhisperX Groq Claude LiteLLM edge-tts Docker

from rageval import track

@track(project="my_rag_app")
async def answer(question): ...
# That is the entire integration.

pip install omnismart-rageval          # core
pip install "omnismart-rageval[eval]"  # + multi-judge scoring and embeddings
pip install "omnismart-rageval[all]"   # everything
rageval init && rageval serve --port 8003

• 5 scoring dimensions: retrieval relevance · groundedness · faithfulness · cost · latency
• Multi-judge consensus — Claude Haiku 4.5 + Groq LLaMA 3.3 + GPT-5-mini; disagreement triggers human-review flag
• Persona-aware: catches when a CFO response pulls data outside its Finance scope
• SQLite by default — zero infrastructure. Postgres + pgvector optional. OpenTelemetry export.
• 12 endpoints, 38 tests, AGPL-3.0

FastAPI sentence-transformers LiteLLM SQLite/Postgres OpenTelemetry React

Multi-source ingestion with domain auto-classification and a live classified dashboard.

• 6 source types: JSON, CSV, Gmail email, webhooks (HMAC-verified), Google Sheets, custom n8n nodes
• Hybrid classifier: keyword fast-path (zero cost) → BGE embedding fallback → Claude Haiku zero-shot
• /webhook/{source}/with-vision — composes with DocIntel /classify-image for auction/inventory aggregation
• First-class n8n integration: custom node template + 3 importable workflows (auction_aggregator, invoice_intake, crm_sync)
• Prefect 3 orchestration for retried, scheduled runs · dlt declarative sources for Gmail/Sheets/webhook
• Live dashboard via WebSocket + SSE · SQLite or Postgres store
• 12 endpoints, 35 tests, AGPL-3.0

FastAPI PostgreSQL pgvector DuckDB React n8n Prefect 3 LiteLLM

LiteLLM             Multi-provider LLM router (no vendor lock-in)
Claude Sonnet 4.6   Reasoning tier (executive personas, deep analysis, sales call)
Claude Haiku 4.5    Judge tier (classification, scoring, support)
Groq LLaMA 3.3 70B  Speed tier (high-volume RAG, meeting notes, general)
Ollama              Local tier (Llama 3.3, Qwen2.5-VL 7B, Llama 3.2 Vision)
OpenAI              Voice agent (VoiceFlow /realtime) + RAGeval 3rd judge (gpt-5-mini)
Gemini              VoiceFlow /realtime fallback when only GEMINI_API_KEY is set
BGE-large-en-v1.5   Dense embeddings
BGE Reranker v2-m3  Cross-encoder reranking
BM25 + RRF          Sparse retrieval + fusion
FastAPI + Uvicorn   API layer across all 6 projects
PostgreSQL + pgvector  Production data store
Qdrant              Vector store in prod (Chroma in dev)
LangGraph           Multi-agent workflows (AgentKit)
FastMCP             MCP server framework (AgentKit)
React + Recharts    Frontends (all 6 have a live dashboard at /)
Docker              Containerized deployment
WhisperX            Local speech transcription (VoiceFlow)
Tesseract           OCR fallback (DocIntel)
Prefect 3           Pipeline orchestration (StreamPulse)
dlt                 Declarative sources (StreamPulse)
PyTorch             Sentence-transformers backend
OpenTelemetry       Observability export (RAGeval)
Tavily              Real-time web search with citations (IntelAI)

• Software Engineer — HyperTech Niger (2025): Designed and deployed full-stack AI and IoT systems, including an ML-driven smart irrigation engine and an energy management platform.
• B.Sc. Artificial Intelligence — African Development University (2022–2025)
• Associate Data Scientist — Qwasar Silicon Valley (2023–2024), trained to SV standards across 57 projects
• IBM certifications: Full Stack Developer · AI Engineering · RAG & Agentic AI · Data Science
• Languages: French (native) · English (fluent) · 🇳🇪 Zarma / 🇳🇪 Hausa (native)

• Upwork: Hire me
• LinkedIn: Yacine Seybou Siddo
• Email: contact@ysiddo-ai-projects.app