Selected work.

अक्षराNet — Sanskrit Devanagari OCR

A segmentation-free deep-learning OCR system that reads printed Sanskrit in the Devanagari script and returns editable Unicode text.

Objective

Recognise printed Devanagari without character segmentation — the shirorekha, conjuncts, and stacked matras that break classical OCR are learned implicitly instead.

My role

Sole author — synthetic data generator (HarfBuzz/FreeType rendering across 7 fonts + scan degradations), CRNN architecture, CTC training loop, full-page line segmentation, FastAPI backend, and a Flutter web front end.

Result

End-to-end pipeline working: pages segment to lines, lines decode to valid Unicode, the API serves real predictions. Isolated-glyph CNN at 99.85% test accuracy; line-level CRNN at 14% CER on held-out synthetic data — a text-limited corpus, not a broken architecture, and documented honestly.

• PyTorch
• CRNN + CTC
• HarfBuzz
• FastAPI
• Flutter

Indian Equity Price Forecasting — LSTM vs. Baselines

A controlled benchmark of an LSTM model against simple statistical baselines for next-day price prediction across NSE equities.

Objective

Test whether an LSTM meaningfully beats Naive, Linear Regression, and ARIMA baselines on next-day price forecasting for 40 NSE equities.

My role

Sole author — data pipeline, feature scaling, model training and evaluation, and the comparative write-up.

Result

The LSTM matched baseline directional accuracy (~50%) but produced 3–4× higher error, which I traced to scaler extrapolation on out-of-range values.

• Python
• TensorFlow
• scikit-learn
• Streamlit

Personal Ride-Data Analysis Dashboard

A Power BI dashboard built over exported personal Uber ride history to surface travel and spending patterns.

Objective

Turn raw ride history into a readable view of when, where, and how much I travel.

My role

Sole author — data cleanup in Excel, modelling, and dashboard design in Power BI.

• Power BI
• Excel

IMDb Sentiment Analysis — DistilBERT

An end-to-end NLP application that classifies movie reviews as positive or negative and returns a confidence score in real time.

Objective

Fine-tune and deploy a production-ready sentiment classifier while keeping inference fast enough to run within free-tier hosting constraints.

My role

Sole author — IMDb data preparation, DistilBERT fine-tuning and evaluation, INT8 ONNX optimization, FastAPI inference API, React interface, and deployment across Render and Vercel.

Result

Reached 93.17% test accuracy and reduced the inference model from a 256 MB FP32 ONNX export to a 64 MB INT8 artifact, enabling a live deployment with confidence-scored predictions.

• DistilBERT
• PyTorch
• ONNX Runtime
• FastAPI
• React
• Vercel

Video2Context — Screen Recordings as Evidence for Coding Agents

An open-source command-line tool that turns a screen recording into timestamped screenshots, speech and on-screen text that Claude Code, Codex or Cursor can read and act on.

Objective

Let a developer explain a bug or UI feedback by recording their screen, then hand the agent only the frames that actually changed — with what was said and what was on screen at each moment — instead of a video it cannot watch.

My role

Sole author — single-pass FFmpeg sampling with NumPy change scoring and perceptual-hash dedupe, local faster-whisper transcription, Tesseract OCR, speech-to-frame fusion, atomic Markdown + JSON package export, a doctor command for setup checks, and a CI/release pipeline backed by 57 tests.

Result

Turns a 2.5-minute macOS recording into 36 screenshots and a context file in about 9 seconds, entirely on-device with no API keys. Replacing per-frame decoder seeks with a timestamp-based FFmpeg pass fixed a 20-second drift on variable-frame-rate QuickTime captures and more than halved sampling time.

• Python
• FFmpeg
• NumPy
• faster-whisper
• Tesseract
• Click