AI Systems Engineer · B.Sc. Artificial Intelligence · Niamey, Niger

I build AI systems that ship, and I publish what I find — including the results that don't flatter the system. My work spans production AI engineering for clients in Niger, six open-source AI/ML tools, and Cibiyar Karatu, an adaptive learning platform for Niger and the Sahel.

Email: contact@yacineseybousiddo.me · LinkedIn · Portfolio Location: Niamey, Niger · Languages: French, English, Hausa, Zarma

• What I Do
• Background
• Open-Source Ecosystem
• Results at a Glance
• Client Work
• Research Direction

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

AI systems engineering. Production RAG, multi-agent orchestration, document intelligence, LLMOps, and edge/IoT platforms, delivered as a freelance engineer for clients in Niger — engineered against Sahel-wide infrastructure constraints (intermittent connectivity, shared NAT, low-end hardware), not yet delivered outside Niger.

Open-source research tooling. Six standalone AI/ML tools, dual-licensed (AGPL-3.0 / commercial), each shipped with a public evaluation protocol rather than a marketing claim.

Adaptive learning platform. Cibiyar Karatu, subject-agnostic and offline-first, for Niger and the Sahel — French alongside Hausa, Zarma, and Fulfulde — plus a parallel research contribution in corpus and benchmark work for those three languages, at very different starting points of existing NLP coverage.

Education

• B.Sc., Artificial Intelligence — African Development Universalis, Niamey (2022–2025), average 15.80/20. Coursework: mathematics for AI, algorithms, ML/DL, NLP, computer vision, databases, systems and networks.
• Associate Data Scientist Program — Qwasar Silicon Valley (2023–2024), 57 project-based assignments evaluated through automated grading and peer review, fully in English.

Certifications — IBM Full Stack Software Developer · IBM RAG and Agentic AI · IBM AI Engineering · IBM Data Science Professional (all 2025) · Mathematics for Machine Learning and Data Science, DeepLearning.AI (2023). Full list of 19 on the portfolio.

Teaching — Instructor, Introduction to Artificial Intelligence, A.D.U. (2025) · workshop lead, Git/GitHub, Tech Communities Club A.D.U. · trainer, "Master AI Tools" workshop.

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.

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. Hybrid retrieval (BGE-M3 dense + BM25 + RRF fusion + cross-encoder reranking) plus a GraphRAG-lite layer for multi-hop queries.

Measured results: out-of-sample backtest on 378 forecasts, mean APE 4.64% (median 2.77%), down from a 12.48%/9.90% prior baseline after fixing a curvature/entity-coverage bug; GraphRAG-lite reaches 95.0% entity coverage; a 50-case production sample scores 71.4% ground-truth accuracy with a judge-panel groundedness of 0.572. Published alongside those numbers: French/English answer-quality parity is currently uneven — 0.917 FR vs 0.431 EN.

Extracts structured data from PDFs and images across three routes: hosted vision LLM, local Ollama vision model (GPU), and a Surya-OCR-plus-LLM fallback (Tesseract as its own automatic fallback). Multi-currency and multi-locale normalization, including FCFA/XOF under UEMOA VAT convention.

Measured results: 95.0% zero-shot accuracy on the SROIE benchmark; 100% on a multilingual cloud-route invoice set; the local GPU route reaches 97.8% field accuracy on a 106-document global + French/FCFA sample, at a fraction of the cloud route's per-document cost.

from rageval import track

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

Multi-judge consensus scoring with bootstrap confidence intervals, not point estimates.

Measured results: HaluEval-QA (N=200), accuracy-weighted consensus 0.860, ROC-AUC 0.902. Published as the headline finding: the multi-judge consensus still did not outperform its single best judge — a negative result, kept in the title rather than buried.

Gives Claude Desktop, Cursor, or any LangGraph agent governed access to live business data through the Model Context Protocol — tool calls, stable resource URIs, and a reusable executive-briefing prompt.

Measured results: 14/14 adversarial guardrail tests pass, deterministic and offline; a 43-scenario LangGraph rerun across all 10 KPI domains passes 43/43; a live rerun against the running MCP server scores 12/12 (100%) on tool-selection accuracy across 10 enterprise domains.

Multi-source ingestion with n8n automation integration and a classification cascade tuned against its own score distribution rather than a fixed threshold.

Measured results: cascade accuracy reaches 99.0% (0.990 macro-F1, N=504) across three stages on held-out telemetry events. Sustained multi-worker ingestion reaches 46.8 requests/second with a 0.0% error rate under burst load.

Routes recorded audio to per-analysis-type LLMs (meeting, sales call, support call, interview) with a multi-provider transcription and diarization fallback chain.

Measured results: 2.2% WER / 0.8% CER on LibriSpeech test-clean (Whisper large-v3, N=150), verified on N>=500, with WebSocket connection latency under 1.8s (1.157s mean).

HyperTech Electronics — AI layer for an e-commerce and retail-operations platform: bilingual semantic search, a six-persona intent-routed assistant powered by RAG, two recommendation engines and a deterministic degraded-mode fallback built after a live provider-quota outage in production.

HyperTech Connect — Designed a zero-trust, protocol-agnostic IoT and edge-management platform for low-connectivity environments, with secure edge networking, multi-protocol device integration, and resilient OTA update capabilities. Validated the platform at 1,000+ devices under test and conducted physical validation on embedded hardware.

HyperFlow — digital-agriculture platform for smart irrigation in Niger, built for Sahel-wide infrastructure conditions, including migration tooling for legacy, informally-structured land-ownership records.

Cibiyar Karatu ("Centre of Learning" in Hausa) is a subject-agnostic, offline-first adaptive learning platform for Niger and the Sahel, spanning formal, non-formal, and informal learning — French instruction alongside Hausa, Zarma, and Fulfulde, on entry-level Android hardware, with or without a network connection. Alongside the platform, it produces a parallel research contribution: corpus and benchmark work for Hausa, Zarma, and Fulfulde — three languages with wide disparity in existing NLP coverage despite their combined speaker count. Zarma and Fulfulde remain acutely under-resourced; Hausa, though still classified as low-resource, already has meaningful benchmark and corpus infrastructure this work builds on rather than starts from scratch. Currently in prototyping and testing, pre-incorporation; architecture and product design are not detailed publicly.

I'm also working toward graduate-level engineering training, aimed at closing the hardware and robotics gap my own software work keeps surfacing — physical AI: natural-scene computer vision, robotics, and embedded systems, beyond the document- and simulation-bound versions I've shipped so far.

Full evaluation protocols and results for each tool are documented in their respective repositories.