# Yaçine Seybou Siddo
> AI Systems Engineer · Independent Researcher · Educator · Entrepreneur
> I turn complex problems into working AI systems — and I teach and research what I build. Based in Niger, open to collaborations worldwide, working at the intersection of low-resource languages, physical AI, and technology for good.
> Portfolio: https://yacineseybousiddo.me
> GitHub: https://github.com/Yacine-ai-tech
> Contact: contact@yacineseybousiddo.me

## Profile
I’m an independent AI systems engineer, researcher, educator, and entrepreneur based in Niger — open to collaborations and opportunities worldwide. My work spans production-grade AI systems deployed in real conditions, research toward trustworthy AI for low-resource languages and physical agents, and teaching across AI, IoT, robotics, and mathematics. I’ve earned certifications from IBM, Duke, Meta, and Google, and I’m building my first venture, Cibiyar Karatu, around applied AI for the Sahel and beyond. My interests span healthcare, agriculture, education, security, and justice — domains where AI can matter most, anywhere in the world.

## Published systems
- IntelAI | Designed and engineered IntelAI, a persona-aware executive analytics copilot combining enterprise KPI intelligence, hybrid retrieval, graph-based reasoning, and time-series forecasting. Supports 9 C-suite personas, 146 KPIs across 7 domains, and 78 months of historical data. Achieved 100% Qdrant retrieval precision, 100% entity coverage across 7,878 rows, 8/8 successful multi-hop queries, and 4.64% MAPE across 444 out-of-sample forecasting backtests compared with a 12.48% baseline. Published as a Python package with technical documentation, benchmark protocols, and empirical research findings. | https://github.com/Yacine-ai-tech/IntelAI
- Cibiyar Karatu | Subject-agnostic, offline-first adaptive learning platform for Niger and the Sahel, spanning formal, non-formal, and informal learning in French alongside Hausa, Zarma, and Fulfulde on entry-level Android hardware, with or without a network connection. Currently in prototyping and testing; real-user testing is upcoming. | 
- DocIntel | Engineered a multimodal document intelligence system through three processing routes, balancing accuracy, deployment flexibility, and processing cost. Native vision, self-hosted vision, and OCR-assisted routes reached 95.0%, 97.8%, and 96.3% accuracy respectively, with 100% throughput reliability across a 650-document multi-format evaluation corpus. | https://github.com/Yacine-ai-tech/DocIntel
- RAGeval | Designed a self-hosted LLMOps evaluation and observability framework for tracking, caching, and assessing LLM outputs through heterogeneous judge consensus and deterministic symbolic verification. Symbolic verification improved consensus ROC-AUC from 0.8709 to 0.9191 on an evaluated HaluEval-QA sample. | https://github.com/Yacine-ai-tech/RAGeval
- AgentKit | Built a governed Model Context Protocol (MCP) tool server and agent-engineering framework focused on reliable tool selection, execution, deterministic guardrails, and multi-agent orchestration. Achieved 100% tool-selection accuracy across 12/12 tools, passed 16/16 adversarial guardrail checks, and completed 43/43 multi-agent scenarios. | https://github.com/Yacine-ai-tech/AgentKit
- StreamPulse | Engineered a business event-processing pipeline combining multi-tier classification, bidirectional n8n webhook workflows, and secure event ingestion. It reached 99.0% accuracy on 504 held-out events, sustained 46.8 requests per second with 0.00% errors, and reached 89.1 requests per second in the evaluated two-instance configuration. | https://github.com/Yacine-ai-tech/StreamPulse
- VoiceFlow | Developed a speech-to-structured-intelligence and voice-agent system combining speech recognition, real-time bidirectional audio streaming, structured action-item extraction, and AI-agent tool integration. Whisper large-v3 reached 2.2% WER on 150 LibriSpeech samples, with 92.4% action-item precision and sub-two-second evaluated handshakes. | https://github.com/Yacine-ai-tech/VoiceFlow
- AI-Powered E-Commerce & Retail Operations Platform | Designed, developed, and supported an e-commerce and retail-operations platform tailored to the Niger/Sahel market. The platform combines online shopping, B2B commerce, inventory management, retail integration, and AI-powered customer experiences. | 
- Secure IoT & Edge Device Management Platform | Designed and developed an IoT and edge-management platform for distributed devices in low-connectivity, infrastructure-constrained environments. It supports secure remote management, reliable telemetry, efficient firmware updates, and scalable operations across industrial monitoring, agriculture, and connected infrastructure. | 
- HyperFlow | Built the software layer of HyperFlow, the smart irrigation system, including data ingestion, analysis, recommendations, exporting, scheduling and control interfaces, and AI-based recommendations. | 
- Energy-Sector SME | Designed the AI layer for an energy-sector SME engagement: consumption forecasting, cost optimization, and NLP on energy-policy documents. | 
- Corporate Web Platform | Rebuilt a corporate site and six subsidiary sites with centralized administration. | 
- DermatoCare Niger | Designed and developed a cross-platform clinical management workstation for dermatology practice in Niger, integrating patient records, consultations, appointment scheduling, inventory, and analytics. The offline-first system includes encrypted records, secure transfers, verified backups, offline French OCR with clinical review, and opt-in local AI with audit trails and fairness controls. | 
- Ilimi App | A QR-code ticketing and check-in application designed and deployed for the 2nd edition of TEDx A.D.U., handling real-world event check-in and ticketing. | 
- IMDB API & Allociné | A Flask-based JSON API serving IMDB movie data, paired with advanced SQL join queries against an Allociné dataset. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley
- First Scraper & DS Babel | A web scraper extracting repository links from GitHub's trending page, paired with a geological data pipeline (fault-line and volcano datasets) converted into queryable SQL databases. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley
- My Mastermind / My printf | Two low-level C exercises from Qwasar's curriculum: a full implementation of the Mastermind game (4-peg color guessing with placement feedback, extended past the base spec) and a custom reimplementation of the printf() function handling multiple format conversions. Both validated through Qwasar's automated test suite. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley/tree/main/Season%201%20Arc%201
- Vivino | Scraped wine data directly from Vivino, then built a recommendation and price-prediction model from wine characteristics and aggregated user ratings. | https://github.com/Yacine-ai-tech/My-data-science-projects/tree/main/My%20Vivino
- Open the Iris | Classic ML classification pipeline on the Iris dataset: data visualization followed by six algorithms (Decision Tree, Gaussian Naive Bayes, k-NN, Logistic Regression, LDA, SVC) compared via cross-validation. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley/tree/main/Season%202%20Data%20Science/my_open_the_iris-dev/my_open_the_iris
- Tu Verra | Data-cleaning and correlation analysis on a housing dataset, followed by a linear regression model predicting house price from the most correlated features. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley/tree/main/Season%202%20Data%20Science/my_tu_verras-dev/my_tu_verras
- My Mob App Studio | Market analysis of a mobile-app dataset — categories, install trends, and pricing — to derive insights for a hypothetical new app's development and go-to-market strategy. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley/tree/main/Season%202%20Data%20Science/my_mobapp_studio-dev/my_mobapp_studio
- NBA Game Analysis | Parses raw NBA game data using regex-based pattern matching to extract individual player statistics and compute team totals. | https://github.com/Yacine-ai-tech/Qwasar-Silicon-Valley/tree/main/Season%202%20Data%20Science/my-nba-game-analysis-dev/my-nba-game-analysis
- M & A | A data pipeline that merges multiple CSV sources into a single structured SQLite database, handling header consistency and data integrity across files. | https://github.com/Yacine-ai-tech/My-data-science-projects/tree/main/My%20M%20And%20A

## Published claims
- Adversarial guardrail cases: 16/16 enforced correctly, reproducible deterministically with zero network or LLM calls  | dataset= | N= | verified=true
- SROIE public benchmark: 95.0% (57/60) zero-shot  | dataset= | N= | verified=true
- Route A accuracy: 92.5-100%  | dataset= | N= | verified=true
- Cost: $0.0007-0.0021/doc (Route B) vs $0.0048-0.0122/doc (Route A)  | dataset= | N= | verified=true
- Route B accuracy (self-hosted Ollama Qwen 2.5-VL): 97.8% global sample (106 documents), 100% French/FCFA sample (325/325 fields)  | dataset= | N= | verified=true
- Throughput/reliability: 650/650 documents processed successfully (100%) at ~1.1 docs/second  | dataset= | N= | verified=true
- Forecasting backtest: Mean APE 4.64% (median 2.77%); auto-selection reduces error from 12.48% (linear-only) to 4.64%  | dataset= | N= | verified=true
- A/B test: Confirmed persona-based RBAC actually filters retrieval, not just display  | dataset= | N= | verified=true
- Production RAG accuracy: 71.4% ground-truth accuracy (20/28 applicable cases), 50/50 cases completed with 0 crashes  | dataset= | N= | verified=true
- GraphRAG-lite entity coverage: 100.0% coverage (7,878/7,878 rows), 8/8 multi-hop queries  | dataset= | N= | verified=true
- MCP tool-invocation benchmark: 12/12 (100%) tool-selection accuracy, 12/12 execution success, Avg. Tool Execution Time 28.0 s, P95 49.5 s, P95 single-tool call latency 1.12 s  | dataset= | N= | verified=true
- 4-judge heterogeneous consensus (HaluEval-QA): 0.860 accuracy, 0.867 F1, 0.902 ROC-AUC (N=200, weighted panel: Claude Haiku, Groq gpt-oss-120b, GPT-5-mini, Gemini 3.5 Flash)  | dataset= | N= | verified=true
- Panel disagreement signal: Stdev 0.217 on wrong predictions vs 0.082 on correct ones (0.158 vs 0.026 with variance penalty)  | dataset= | N= | verified=true
- Webhook HMAC verification: 100% correct accept/reject (90/90 valid processed, 10/10 invalid rejected)  | dataset= | N= | verified=true
- WER/CER: 2.2% WER, 0.8% CER (LibriSpeech test-clean, Whisper large-v3, N=150, T4 GPU)  | dataset= | N= | verified=true
- Classification cascade accuracy: 8.3% (keyword only) -> 64.6% (+embedding) -> 99.0% (full cascade), macro-F1 0.990 (N=504 held-out SaaS telemetry events)  | dataset= | N=504 | verified=true
- Action-item and sentiment extraction: 0.812 precision, 0.729 recall, 0.763 F1 (gpt-oss-120b, N=50), 0.759 precision, 0.765 recall, 0.755 F1 (claude-sonnet-4-6, N=50)  | dataset= | N= | verified=true

## Published deployments
- IoT Network & Edge Gateway | IoT Network & Edge Gateway | {} | Production / Validated
- Digital Agriculture & IoT | Digital Agriculture & IoT | {} | Field Deployed
- Hybrid Commerce & Operations | Hybrid Commerce & Operations | {} | Operational in Production
- DermatoCare Niger | Clinical workflow platform | Deployed cross-platform workstation for patient records, consultations, appointments, inventory, analytics, offline operation, and responsible AI workflows. | Live — client deployment
- Industrial Energy & AI | Industrial Energy & AI | {} | Delivered & Validated
- IntelAI | Enterprise Analytics Copilot & GraphRAG-lite | Backtested forecasting error: 4.64% mean MAPE across 444 out-of-sample 3-month forecasts over 6 KPIs | Live — Public Demo API
- DocIntel | Zero-Shot Document Extraction (Multi-Route OCR/LLM) | 650/650 documents processed at 100% success (~1.1 doc/sec); 97.8%-100% field accuracy on the self-hosted route | Live — Public Demo API
- RAGeval | RAG Observability & Multi-Judge Groundedness Scoring | 4-judge weighted consensus: 0.860 accuracy / 0.902 ROC-AUC on HaluEval-QA (N=200) | Live — Public Demo API
- AgentKit | Governed MCP Multi-Agent Orchestration | 43/43 (100%) autonomous 3-node LangGraph runs across 10 KPI domains; 14/14 adversarial guardrails enforced | Live — Public Demo API
- StreamPulse | Real-Time Telemetry Classification Pipeline | 99.0% macro-F1 classification cascade (N=504); 46.8 req/s sustained burst ingestion at 0.00% error | Live — Public Demo API
- VoiceFlow | Speech Intelligence & Meeting-Notes Extraction | 2.2% WER / 0.8% CER on LibriSpeech test-clean (GPU); 1.157s mean latency on Gemini Live streaming | Live — Public Demo API

## Published credentials
- IBM AI Engineering Professional Certificate (V3) | IBM | Dec 2025
- IBM RAG and Agentic AI Professional Certificate | IBM | Dec 2025
- IBM Full Stack Software Developer Professional Certificate (V5) | IBM | Dec 2025
- Mathematics for Machine Learning and Data Science | DeepLearning.AI | Nov 2023
- IBM Data Science Professional Certificate | IBM | Jul 2024
- IBM Data Science Professional Certificate (V3) | IBM | Oct 2025
- Google IT Support Professional Certificate | Google | 
- Agile Project Management | Google | Feb 2024
- Leading Transformations: Manage Change | Macquarie University | Apr 2024
- AI Product Management Specialization | Duke University | 
- Innovation & Entrepreneurship: From Basics to Open Innovation | 28DIGITAL | Dec 2024
- Understanding Research Methods | University of London (SOAS) | Feb 2024
- Principles of UI/UX Design | Meta | Feb 2024
- Introduction to Data Engineering | IBM | Nov 2024
- Artificial Intelligence in Marketing | University of Virginia | Dec 2024
- Ethical Hacking |  | 
- Associate Data Scientist | Qwasar Silicon Valley | Jul 2024
- Introduction to Software Engineering | Qwasar Silicon Valley | 
- Introduction to Data | Qwasar Silicon Valley | 

## Recent posts

## Published profile metrics
- Open-source systems shipped: 6 | IntelAI, RAGeval, DocIntel, StreamPulse, VoiceFlow, AgentKit — deployed, tested, documented
- Deterministic guardrails: 16/16 | 16/16 adversarial tests passed with no network call and no LLM dependency via AgentKit
- Zero-shot SROIE accuracy: 95.0% | Measured on the public SROIE benchmark (57/60 annotated receipts) via DocIntel
- Research blocks in development: 3+1 | Ancre active; Falaisa and Gargi warming up; Trans-Saharan Engine inside Cibiyar Karatu

## Data provenance
> This document is generated from published PostgreSQL records.