I build systems that turn signal into something you can act on.

A system that listens to speech and works out not just whether the speaker is positive or negative, but whether they mean it. Sarcasm is the hard case: the words and the tone disagree, so reading the transcript alone gets it wrong.

It reads both signals against each other — language through mBERT embeddings and a trained classifier, tone through pitch and energy features from Librosa — then fuses them. Around that sits a full pipeline: Whisper transcribes the audio and detects the language, a router translates across seven languages using NLLB-200 or Marian, and RoBERTa handles sentiment.

It runs as a full application with JWT authentication, bcrypt-hashed passwords, per-user analysis history in MongoDB, an admin panel, and PDF export with embedded Noto fonts so Indic scripts render properly rather than as boxes.

A Power BI report over six months of transaction-level retail sales — 149,000 records across three locations. Built to answer the questions a shop owner actually asks: where revenue is trending, which hours carry the day, which products are doing the work.

The interesting part is not the charts, it is the interaction. Most dashboards answer one question and then need an analyst for the follow-up. This one lets the viewer ask their own: a single slicer switches the trend and hourly charts between four metrics, right-clicking any product opens a detail page that renames itself to whatever was clicked, and hovering shows a miniature report rather than one number.

Underneath it is a star schema with a marked date table, so month-over-month and year-to-date resolve correctly rather than approximately. Fifteen measures, and the whole thing styled through a theme file — nothing in it is a Power BI default.

A Java desktop application that reads sensor data as it arrives and shows temperature, humidity, and air quality in a form someone can scan in a second, with alerts when a reading crosses a threshold.

The real problem was legibility under change — values updating constantly need a layout that holds still, and a schema that stays queryable as historical readings pile up.

The multi-sensor chart is drawn from scratch in Graphics2D rather than pulled from a library, and a Python layer runs alongside the Java client for linear-regression forecasting and z-score anomaly detection.