// power platform & power bi developer
I build business apps, workflow automation and reporting on Microsoft 365, and the Python tools around them. Three years of Power Apps, Power Automate, SharePoint and Power BI, from freelance projects to apps in daily use on construction sites.
Open to junior Power Platform, BI and data roles in Germany and Italy.
Hildesheim, GermanyWorking student at S01 Projektmanagement GmbHM.Sc. Data Analytics, University of Hildesheim
// what I ship with, most days
{
"platform": ["Power Apps", "Power Automate",
"SharePoint", "SPFx", "Copilot Studio"],
"data": ["Power BI", "DAX", "SQL", "Python"],
"ai": ["RAG", "spaCy", "Tesseract OCR", "LLM APIs"],
"tooling": ["Power Platform CLI", "PnP PowerShell",
"Git", "Playwright"],
"languages": { "en": "C2", "it": "A2",
"de": "A1", "ta": "native" },
"since": "2022"
}Working student at S01 Projektmanagement GmbH since July 2025. The website demos are public. The other apps are internal, so they are shown here as clickable demos under a fictional company name with invented data.
Six fully navigable redesign demos of s01-pm.de in German and English, generated from one bilingual content source by a Node static generator with a component kit and a motion system. Unit tests and an output checker validate links, translations and colour contrast on every push to main; Playwright scripts measure the layout at up to seven viewport widths.
Offline-capable Power Apps canvas app for site inspectors, covering quality inspections and schedule tracking: photo capture, video upload, automatic SharePoint sync, and 45 app revisions shaped by supervisor feedback. Behind it: a SharePoint data model of 9 lists and a document library, and 11 Power Automate flows for observation numbering, photo and video filing, Excel imports, route export and Word reports e-mailed to a chosen recipient.
A full CRM built with SPFx in TypeScript on SharePoint lists: leads pipeline, companies, contacts, tasks, reports (hit ratio, activity, quarter results) and OKR scoring, lead creation from a pasted email, CSV, Excel-paste and vCard import, CSV export, English and German interface. Versioned releases that the supervisor deploys by dropping one file into a library.
Portable, offline Windows app that permanently removes client names, personal data, logos and signatures from PDF, Word, PowerPoint and Excel files before they leave the company: signature lines automatically, signature images on request. Automatic client-name detection, German NLP for person names, an interactive preview to keep or redact single spots, and a four-part quality check, including OCR, that marks every output PASS or FAIL. 106 automated tests.
A Copilot Studio assistant that answers inspectors' questions on technical standards and cites chapter and page, built on a self-made chapter index of the standard documents. Compared five routes (Copilot Studio, Microsoft 365 Copilot with a SharePoint agent, Azure AI Foundry with Azure OpenAI and AI Search, the Claude API, and AI Builder) against licence and data-residency limits and wrote the decision paper. Also repaired an existing invoice-archiving flow in Power Automate and fixed the root cause of its failures.
Two take-home assignment solutions presented as one interactive page: a sales CRM on SharePoint lists with Power Apps screens, and a construction-site QA app that captures photo, defect and comment and generates a structured report. Both are playable inside the presentation.
Source on GitHub, live where a browser is enough.
Electron app on top of the open-source Career Ops engine: profiles with isolated workspaces, pipeline, discovery, CV intake, agent sessions with the Claude or Codex CLI, practice interviews and five themes. React and Vite renderer, Electron main process, 896 unit tests.
Six small machine-learning demos that run entirely client-side, each on its own page with the interactive part next to the maths: TF-IDF retrieval for RAG, a transformer attention heatmap, a LoRA parameter and memory calculator, an optimiser race on a loss surface, step-by-step k-means, and a naive Bayes sentiment classifier.
Upload a PDF passport, ID card or driving licence and get surname, given name, dates and document number extracted into an editable form. Express API with multer uploads and regex extraction over the PDF text layer; React and TypeScript front end with drag-and-drop.
Desktop tool that recognises known faces from a webcam with OpenCV and dlib, marks attendance with timestamps and exports CSV, with a Tkinter interface and a dlib-free simple version for machines where dlib will not build.
Text-to-SQL benchmark in German and English on a SQL Server energy data warehouse. 70 reference questions with gold SQL (drafted by a language model, verified by execution), a fail-closed SQL guard with a read-only login, six prompt variants, scoring with an error taxonomy, Anthropic and Ollama adapters and a DuckDB sample in CI. Live model runs are still pending; until then the results page shows only the pipeline test.
Probabilistic day-ahead forecast of German grid load (SMARD, CC BY 4.0) with LightGBM quantile regression and conformal calibration. Pre-registered backtest protocol (tag protocol-v1) over 730 test days: CRPS 42 % below the seasonal-naive baseline, with 80 % and 90 % intervals covering 80.1 % and 89.8 %. FastAPI service in a Docker image and a daily live run into a forecast-log branch. Tag v0.1.0 is still to come.
Evaluation bench for retrieval-augmented generation on five German energy laws (public domain, § 5 UrhG). 240 questions drafted by a language model, human review still pending; six retrieval variants times two chunkings, scored with Recall@k, MRR and nDCG with bootstrap intervals. In the published run, hybrid plus reranker reaches Recall@5 of 0.71 against 0.54 for BM25, and the gain comes from the reranking. LLM layer with Anthropic and Ollama backends and record and replay. Results page and tag v0.1.0 follow; answer and judge runs are pending.
Analytics-engineering pipeline for Italian public procurement data (ANAC Open Data, CC BY-SA 4.0): an immutable raw archive with checksums, a landing layer that redacts personal data of natural persons before writing, dbt layers with 40 models and contract tests, a star schema and nine marts for a later Power BI report, a data quality report and Dagster orchestration. The dbt docs are on GitHub Pages. The full 2024 to 2026 run and tag v0.1.0 follow.
My thesis asks whether the scope of positive selection in SelfPatch's patch-level loss changes the label-free object masks that a spectral decoder reads out of a self-supervised Vision Transformer. To understand the method before writing it up, I built an interactive course on the two papers behind it and on the thesis's own rule.
A from-zero course for a reader who has never met a neural network: numbers and vectors, the Vision Transformer, DINO's student and teacher, SelfPatch's patch-level loss, and the contrastive selection rule my thesis studies. Every equation is walked with numbers you can check, every design decision gets a "why" card, and the labs run on real photographs with features from DINO, the released SelfPatch model and my own training runs. Four marks keep facts, real computations, toy simplifications and readings apart.
07/2025 — present
S01 Projektmanagement GmbH, Germany
02/2022 — 05/2025
Upwork and Fiverr, remote, international clients
10/2022 — present
University of Hildesheim, Germany
12/2018 — 06/2022
Riga Technical University, Latvia
Based in Hildesheim, Germany, open to relocation. Email is the fastest way to reach me; LinkedIn works too.
harikaranharithas@gmail.com