Skip to content
View Som0111's full-sized avatar
💭
Enjoying
💭
Enjoying

Block or report Som0111

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Som0111/README.md

Hi, I'm Soumya Padhi 👋

B.Tech in Electrical Engineering

National Institute of Technology Rourkela

Machine Learning Data Engineering AI

Building systems that go beyond the notebook

LinkedIn GitHub


I enjoy working at the intersection of Machine Learning, Data Engineering, and real-world engineering problems — with a focus on building systems that are measurable, deployable, and useful beyond a notebook.

🧠 Flagship Projects

🛰️ Contrail · Event-Time Stream Processor for Flight Telemetry

Python · FastAPI · Kafka (Redpanda) · TimescaleDB · Redis · Docker · Prometheus/Grafana

Contrail questions

Contrail ingests live and synthetic ADS-B flight telemetry through Kafka and processes it with correct event-time semantics — the three questions that separate a real streaming system from a Kafka-consumer tutorial: how do you handle data that arrives late, how do you scale under burst load, and how do you guarantee correctness after a crash.

Event misattribution (naive → watermark) Peak lag under 6x burst Replay determinism
13.66% → 0.00% 4.2x lower (adaptive + shedding vs. static) Byte-identical across repeated runs and mid-stream crashes

Every claim benchmarked against a naive baseline, on real ADS-B data and controlled synthetic chaos (out-of-order, duplicate, and late events) — full reproduce commands and exact configs in BENCHMARKS.md.

View Contrail

🔍 How Contrail works
Live OpenSky feed / synthetic generator
   ↓
Kafka (Redpanda) — partitioned, idempotent producers
   ↓
Watermark-based event-time windowing (handles out-of-order + late data)
   ↓
Lag-trend adaptive controller (scales workers, sheds load under sustained overload)
   ↓
TimescaleDB (durable) + Redis (cache/pub-sub)
   ↓
FastAPI — REST + WebSocket, JWT-authenticated
   ↓
Prometheus/Grafana observability
🧪 Engineering highlights
  • Watermark-based windowing with a configurable allowed-lateness bound and a late-event side output, benchmarked against a naive processing-time baseline on 88,199 events.
  • Lag-trend adaptive control loop that scales and sheds Kafka consumer workers based on the rate of change in lag, not a static threshold — cuts peak lag 4.2x and p99 latency ~6x vs. a static pool during a 6x burst.
  • Deterministic replay harness proving byte-identical output across repeated runs and mid-stream consumer failures, backed by idempotent database writes.
  • FastAPI service with JWT auth, rate limiting, REST + WebSocket APIs, live OpenSky ingestion, full Prometheus/Grafana observability, and a chaos-tested crash-recovery path — all runnable with one docker compose up.
  • Documented, honest limitations (not overclaimed): live multi-partition watermarking, shedding-under-sustained-overload — see DESIGN_DECISIONS.md.

🧠 CI-Brain · Continuous Integration Intelligence Platform

FastAPI · PostgreSQL · SQLAlchemy · Alembic · React + TypeScript · Docker · GitHub Actions · Google Gemini · Render

CI-Brain questions

CI-Brain analyzes CI pipelines to make software testing faster and more intelligent — covering the three core problems large engineering orgs actually solve: which tests to run, which tests are flaky, and what caused a wall of failures.

Tests eliminated Wall-clock time saved Flaky tests caught
57.5–98.9% 9.5–58.2% 4 / 4 · 0 false positives

Benchmarked against toolz, a real open-source Python library, across a range of simulated single-file and multi-file diffs — ranges reflect diff size, not a single run.

View CI-Brain Live API

🔍 How CI-Brain works
CI data
   ↓
Coverage / file-to-test dependency graph
   ↓
Test Impact Analysis
   ↓
Statistical Flaky-Test Detection
   ↓
Failure Clustering
   ↓
Evidence-grounded LLM Explanation
   ↓
Developer Dashboard
🧪 Engineering highlights
  • Test Impact Analysis falls back to the full suite when confidence is insufficient.
  • Flaky-test detection uses behavioral evidence rather than a simple failure-rate threshold.
  • Failure clustering groups failures by likely root cause before asking the LLM for a summary.
  • 110 tests with 98%+ branch coverage.
  • FastAPI backend, PostgreSQL database, React + TypeScript dashboard, Dockerized deployment and GitHub Actions integration.

🚀 Featured Projects

Build Measure Deploy Improve

India Electricity Demand Analytics

Python · PostgreSQL · SARIMAX · Power BI · GitHub Actions

End-to-end ELT pipeline analyzing India's electricity demand across 65,178+ state-level daily records (2020–2025).

  • PostgreSQL star schema on Neon
  • 12 SQL analyses using CTEs and window functions
  • SARIMAX forecast — 2.49% MAPE
  • Power BI dashboard
  • Daily automated refresh with data-quality validation

Cost-Sensitive Churn Prediction

scikit-learn · FastAPI · Docker · GitHub Actions · Render

Optimizes the decision threshold around the actual economics of a retention campaign, rather than accuracy alone.

Strategy Profit
Blanket campaign −$14,350
ChurnGuard +$16,250
Swing +$30,600

Profit figures use illustrative campaign-cost assumptions ($50 offer cost / $500 customer LTV / 30% save rate) on the IBM Telco dataset — the framework is the point, not these exact dollar values.

🕵️ FraudLens

Credit Card Fraud Detection

XGBoost · Neural Networks · Jupyter

End-to-end fraud detection including temporal drift analysis, model robustness and business cost optimization — not just AUC.

NILM — Non-Intrusive Load Monitoring

Signal Processing · scikit-learn · REFIT

Identifies household appliances from aggregate power consumption by combining Electrical Engineering, signal processing, and ML.

Random Forest Macro F1: 0.56


Electrical Engineering Projects

Core-EE work applying signal processing and circuit analysis to problems that actually show up on a plant floor — condition monitoring and power quality, not just simulation for its own sake.

Bearing Fault Diagnosis via Envelope Spectrum Analysis

Signal Processing · Hilbert Transform · scikit-learn · CWRU Dataset

Diagnoses rolling-element bearing faults from raw vibration by demodulating the structural resonance — physics first, ML as a cross-check, not the method.

  • Rule-based diagnosis (bandpass → Hilbert envelope → defect-frequency matching): 83.2% accuracy, zero false alarms on healthy bearings
  • RandomForest cross-check: 100% test accuracy, confirming the physics-derived features (envelope amplitude at BPFO/BPFI/BSF) carry the signal
  • Kurtogram-based automatic band selection (Antoni, 2006), including an honest negative result: it doesn't fix the hardest fault class, and that failure is itself diagnostic
  • Self-checked against a synthetic signal with a known answer before trusting it on real bearing data

Plant Feeder Harmonic Filter Design (IEEE 519-referenced)

Python · Circuit Simulation · Power Systems

Simulates a 415V industrial feeder with a VFD load to show how a naive capacitor bank for power-factor correction can make harmonic distortion worse — and why a detuned filter is the standard fix.

  • Harmonic generator verified against theory before any simulation runs
  • Naive capacitor bank: PF improves, but THD rises to 9.99% (from 8.36% baseline) due to resonance amplification, even with the resonant order far from the classic 5th/7th harmonic
  • Detuned filter (tuned below the 5th): lower total kVA and lower absolute harmonic current on every line — the real fix, even though THD% alone doesn't show it clearly
  • Pure Python simulation, no MATLAB/Simulink, referenced against IEEE 519 limits

🔨 Currently Building

Active Python OpenCV MediaPipe ONNX Runtime

A modular, plugin-based computer-vision platform for gesture-controlled applications and AI demos.

Current focus:

  • Plugin-based application architecture
  • Shared camera pipeline
  • Computer-vision inference
  • CPU-efficient processing
  • ONNX Runtime integration

🧩 What I Work With

Core

Python · C++ · SQL

Backend / Systems

FastAPI · Kafka (Redpanda) · PostgreSQL · TimescaleDB · Redis · Docker · GitHub Actions · Prometheus/Grafana

ML / AI

scikit-learn · TensorFlow · XGBoost · OpenCV · ONNX

Data / Signal Processing

Pandas · NumPy · SciPy · Hilbert Transform · Power BI

Frontend / Visualization

React · TypeScript · Power BI


📊 GitHub Activity

GitHub stats Top languages

GitHub activity graph


🎓 Education

National Institute of Technology Rourkela B.Tech — Electrical Engineering · Graduating 2028


📌 Current Focus

  • Backend & Distributed Systems Engineering
  • Machine Learning Engineering
  • Data Engineering
  • Computer Vision
  • AI-powered developer tools

Open to SDE, Machine Learning and Data Engineering internships · Placement 2028

LinkedIn · GitHub

Pinned Loading

  1. contrail contrail Public

    Event-time stream processor for flight telemetry — watermarked windowing, lag-driven adaptive scaling, and deterministic replay, benchmarked against a naive baseline.

    Python

  2. ci-brain ci-brain Public

    Test intelligence platform for CI pipelines — flaky test detection, coverage-graph test impact analysis, and LLM-based failure clustering.

    Python 1

  3. gridpulse gridpulse Public

    Automated ELT pipeline analysing India's electricity demand across 65k records (2020–2025). Python → PostgreSQL (Neon) → Power BI. 12 SQL analyses, SARIMAX forecast (2.49% MAPE), daily GitHub Actio…

    Jupyter Notebook 1

  4. churn-guard churn-guard Public

    Cost-sensitive churn prediction. Tunes the decision threshold on retention-campaign profit instead of 0.5, turning a -$14K blanket campaign into +$16K. sklearn → FastAPI.

    Python 1

  5. motor-condition-monitoring motor-condition-monitoring Public

    Bearing fault diagnosis from vibration signals using envelope spectrum analysis (Hilbert transform) — CWRU dataset, physics-based feature extraction, no deep learning

    Python

  6. power-quality-reactive-compensation power-quality-reactive-compensation Public

    Power factor correction and harmonic filter design for a plant feeder with a VFD load — Python simulation, IEEE 519-referenced

    Python