About  ·  Projects  ·  Skills  ·  Education  ·  Activity  ·  Banner  ·  Contact

I'm an AI & automation engineer in training, based in India, and I learn by shipping. Rather than stopping at a notebook, I take ideas all the way to working products: desktop applications, interactive maps, and models that grade their own accuracy.

My work sits at the intersection of applied machine learning, geospatial tooling, and clean software architecture. I care about the part most people skip: proving that a system actually beats a baseline.

Build it, measure it, prove it beats the baseline.

Currently focused on

• Forecast validation: systems that score their own accuracy against baselines
• Applied ML for real problems: weather, routing, and environmental monitoring
• Deep learning: from-scratch implementations through to working pipelines

flowchart LR
    A([Real problem]) --> B[Build a<br/>working version]
    B --> C{{Measure vs.<br/>a baseline}}
    C -->|beats it| D([Ship + document])
    C -->|doesn't| B

    style A fill:#0ea5e9,color:#fff,stroke:none
    style B fill:#6366f1,color:#fff,stroke:none
    style C fill:#8b5cf6,color:#fff,stroke:none
    style D fill:#22c55e,color:#fff,stroke:none

mindmap
  root((Sumit))
    Forecasting
      Multi-model ensembles
      Self-scoring accuracy
      Persistence and IFS baselines
    Geospatial
      Geocoding
      Road-network routing
      Interactive maps
    Machine Learning
      CS50 AI
      Deep Learning from scratch
      Object detection
    Reinforcement Learning
      Offline RL
      Groundwater audit

AI/ML Intern · National Aluminium Company Limited (NALCO) Jun 2025 – Jul 2025 · Damanjodi, Odisha

• Engineered an edge-deployable YOLOv8 PPE detection system trained on 5,000+ industrial images, achieving 82% classification accuracy and reducing manual inspection overhead by ~30%.

Full-Stack Developer Intern · GIET University Jun 2024 – Jul 2024 · Gunpur, Odisha

• Built a full-stack Library Management System (Python/Flask, SQL) featuring book search, checkout, and admin/inventory controls with optimized queries.

The model is about 20% of the work. The other 80% is association, rules, and operations.

Per-worker PPE compliance from the cameras a site already owns. It binds every PPE observation to a tracked worker, applies a configurable policy, and only commits a violation after an N-of-M temporal vote (6 of 10 frames), so one blurry frame never raises an alert. Events are written to a SHA-256 hash-chained audit trail you can verify. Phase 2 is complete, with a Gradio demo, a CLI, and 289 weight-free tests at 96% coverage; the zone engine, audit store, and dashboard are on the roadmap.

Python · YOLOv8 · ByteTrack · OpenCV · Gradio

View repository →  ·  Live site →

One forecast is a guess; an ensemble is evidence.

A desktop app that queries four independent forecast models in parallel (ECMWF IFS, ECMWF AIFS, GFS, and ICON) and reports the median with a confidence spread instead of a single number. It grades its own accuracy against observed conditions and benchmarks against persistence and IFS baselines, so you can see exactly when the ensemble is worth trusting.

Python · PySide6 · Open-Meteo

View repository →

Range anxiety is a data problem, not a car problem.

Enter a location in India and get the nearest EV charging station with a real driving route drawn on an interactive map: an actual road-network path, not a straight line. Includes a 3D landing page, a live demo, and a Python backend that geocodes, routes, and renders standalone HTML maps you can open or share anywhere.

Python · Folium · geopy · OSRM · Leaflet

View repository →  ·  Live demo →

You can't run experiments on an aquifer.

Applies offline reinforcement learning to groundwater monitoring and audit tasks, learning and evaluating policies purely from historical data with no online interaction. It turns logged records into decision-making policy, which is exactly the constraint real environmental monitoring operates under.

Python · Reinforcement Learning · Offline RL

View repository →

More activity  _{(contribution graph and snake)}

The header at the top changes through nine scenes across the day, timed to India Standard Time, by a small GitHub Action. Each scene is a single hand-coded animated SVG.

See all nine scenes  _{(swapped automatically by IST time of day)}

🌄 Dawn 4:30 – 6:30 AM	🌅 Sunrise 6:30 – 9:30 AM	🌤️ Morning 9:30 AM – 12 PM

☀️ Noon 12 – 3 PM	🌞 Afternoon 3 – 5 PM	🌇 Sunset 5 – 6:30 PM

🌆 Dusk 6:30 – 7:30 PM	🌙 Night 7:30 – 11:30 PM	🌌 Midnight 11:30 PM – 4:30 AM

Interested in forecasting, geospatial tooling, or applied ML? Open an issue, star a repo, or reach out. I'm happy to talk shop.