Aditya Mishra | Portfolio

Machine learning · Computer vision · Real-time inference

Machinelearningthatships.

B. Tech, Instrumentation & Control Engineering at Netaji Subhas University of Technology (2023-2027), focused on machine learning, computer vision, and real-time inference systems.

0.00

ev soc mae

beats 4.28–6.5

−0.0%

inference latency

7.7s → 0.86s

0×

throughput

async + parallel

hackathon reach

IEEE NSUT, Chair

Token usage

256.72M all-time

Codex0

Claude0

Gemini0

01 · toolkit

Technical toolkit

PythonSQLMATLABJavaTensorFlowscikit-learnXGBoostOpenCVFastAPIStreamlitTime-Series ForecastingComputer VisionRaspberry PiMQTTGitJupyter

02 · experience

Where I've worked

ML Engineer Intern

Nov 2025-Present

DUSQ (formerly InnerGize)

Cut end-to-end inference latency by 88.9% (7.7 s to 0.86 s) and 9x throughput via parallel and async execution.
Built a 27-keypoint placement API with visibility-aware training for robust side-view guidance at sub-second speeds.
Optimised preprocessing to 26 ms mean latency (64% faster; P99 37 ms, -66%), then packaged low-latency pipelines for production.

ML-IoT Intern

Jun 2025 - Jul 2025

Indian Meteorological Department

Automated thermometer readings using a Raspberry Pi vision system and OpenCV pipeline.
Trained models on a custom dataset of 1000+ images captured across diverse weather conditions.
Implemented MQTT-based real-time transmission enabling low-latency environmental monitoring.

03 · moments

Life & leadership

With the IEEE NSUT Student Branch team for 2025-26.

Projects

04 · builds

★ Featured · Research

Dec 2025 - Present

Res-PhyT: Physics-Informed Transformer for Battery SoC

Nernst/Arrhenius physics · 4 real-world datasets

Built a Residual Physics-Informed Transformer (encoder-decoder) for EV battery State-of-Charge forecasting, embedding Nernst and Arrhenius physics into a 17-feature preprocessor.
Trained and evaluated across four real-world battery datasets (NASA, Oxford, BIT, TUM) for cross-domain robustness.
Engineered a modular PyTorch pipeline with Hydra configs, Weights & Biases tracking, checkpointing, and a probabilistic head for uncertainty estimates.

TransformerPINNPyTorch

GitHub

02 · Full-stack

Mar 2026

ArthaSutra: Gamified Finance-Learning App

Next.js + Supabase · live on Vercel

Built a gamified finance-learning platform — virtual investing, league progression, and the Aarohan contest.
Engineered dual-mode persistence: Supabase auth + cloud sync, with a local-storage demo fallback when unconfigured.

Next.jsSupabaseRecharts

Live GitHub

03 · Research · Published

May 2024 - Present

EV Battery State-of-Charge Forecasting & Real-Time Inference

MAE 3.66 — beats prior benchmarks (4.28–6.5)

Developed a Bayesian-optimized LSTM model for EV battery SoC prediction using 70 real-world driving sessions.
Achieved MAE of 3.66, outperforming prior research benchmarks (4.28-6.5).
Engineered a real-time inference pipeline using pre-trained Keras models for continuous SoC monitoring.
Integrated SHAP-based explainability to analyze feature contributions in battery behavior.

LSTMBayesian OptimizationKeras

GitHub

04 · Recommender Systems

Mar 2026

CSAO: Cart-Aware Add-On Recommender

NDCG@10 0.707 · p95 latency <200ms

Designed a real-time food-delivery recommender — LightGBM LambdaRank retrieval, a neural cross-attention reranker, and LLM-generated explanations — served under 200ms.
Reached NDCG@10 of 0.707 (+15% vs popularity, +39% vs co-occurrence) with bootstrap 95% confidence intervals and Optuna hyperparameter tuning.
Validated cold-start (unseen user, restaurant, item) and catalog coverage on 952K synthetic order-items across 8 cuisines.

LightGBMLambdaRankLLM

GitHub

05 · Computer Vision · Production

Jun 2025 - Present

CAPTCHArd: CAPTCHA Recognition Service

OpenCV + TF · CI + signed model releases

Built a production Python library to segment and recognize 5-digit CAPTCHAs with OpenCV and TensorFlow, exposing both sync and async prediction APIs.
Hardened delivery with pytest and GitHub Actions CI, uv-locked environments, and SHA256-verified release-based model artifacts.

OpenCVTensorFlowCI/CD

GitHub

06 · Full-stack ML

StockVision: Real-Time Stock Forecasting Platform

LSTM · FastAPI + Streamlit + SHAP

Built end-to-end LSTM forecasting system using technical candlestick features.
Deployed Streamlit + FastAPI dashboard with SHAP-based model explainability.

LSTMFastAPIStreamlit

05 · research

Publication

published

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries

IEEE SEFET 2025 · IEEE Xplore

06 · recognition

Awards & certifications

IEEE Outstanding Student Volunteer Award for Academic & Extracurricular Excellence (June 2026)
IBM Data Science Professional Certificate (2025)
Lala Ram Kishan Das Jain Memorial Award for Initiative (2022)

07 · writing

Thoughts & perspectives

all posts

March 24, 2026

The Harness Problem: Why the Same Model Feels Like a Genius in One Tool and an Intern in Another

Claude on Cursor and Claude on Claude.ai are the same model — but they feel completely different. The gap isn't the AI. It's the harness. Here's what that means, with real Terminal-Bench numbers showing an 18-point swing from scaffolding alone.

read article

March 16, 2026

Bias vs Variance: The Invisible Balance Behind Every ML Model

A clear and engaging explanation of bias, variance, underfitting, overfitting, and why building good machine learning models is really about finding the right balance.

read article

March 14, 2026

Build the MVP Before You Build the Dream

A powerful reflection on why the first version of a product matters more than the perfect version that never reaches users.

read article

March 10, 2026

A Holiday Accident That Quietly Broke Machine Learning's Oldest Rule

OpenAI researchers accidentally left a model training over the holidays. When they came back, a model that had completely failed on unseen data now understood everything perfectly. This is the story of grokking—and why it quietly broke several rules we thought were permanent.

read article

contact

Let's build something intelligent.

Open to ML research collaborations and engineering roles. The fastest way to reach me is email.

aditya_mishra@outlook.in

Email LinkedIn GitHub

Machine learning · Computer vision · Real-time inference

Machinelearningthatships.

B. Tech, Instrumentation & Control Engineering at Netaji Subhas University of Technology (2023-2027), focused on machine learning, computer vision, and real-time inference systems.

0.00

ev soc mae

beats 4.28–6.5

−0.0%

inference latency

7.7s → 0.86s

0×

throughput

async + parallel

hackathon reach

IEEE NSUT, Chair

Token usage

256.72M all-time

Codex0

Claude0

Gemini0

GitHub

@mishra-bytes

last 6 months509 contributions

01 · toolkit

Technical toolkit

PythonSQLMATLABJavaTensorFlowscikit-learnXGBoostOpenCVFastAPIStreamlitTime-Series ForecastingComputer VisionRaspberry PiMQTTGitJupyter

02 · experience

Where I've worked

ML Engineer Intern

Nov 2025-Present

DUSQ (formerly InnerGize)

Cut end-to-end inference latency by 88.9% (7.7 s to 0.86 s) and 9x throughput via parallel and async execution.
Built a 27-keypoint placement API with visibility-aware training for robust side-view guidance at sub-second speeds.
Optimised preprocessing to 26 ms mean latency (64% faster; P99 37 ms, -66%), then packaged low-latency pipelines for production.

ML-IoT Intern

Jun 2025 - Jul 2025

Indian Meteorological Department

Automated thermometer readings using a Raspberry Pi vision system and OpenCV pipeline.
Trained models on a custom dataset of 1000+ images captured across diverse weather conditions.
Implemented MQTT-based real-time transmission enabling low-latency environmental monitoring.

03 · moments

Life & leadership

With the IEEE NSUT Student Branch team for 2025-26.

Projects

04 · builds

★ Featured · Research

Dec 2025 - Present

Res-PhyT: Physics-Informed Transformer for Battery SoC

Nernst/Arrhenius physics · 4 real-world datasets

Built a Residual Physics-Informed Transformer (encoder-decoder) for EV battery State-of-Charge forecasting, embedding Nernst and Arrhenius physics into a 17-feature preprocessor.
Trained and evaluated across four real-world battery datasets (NASA, Oxford, BIT, TUM) for cross-domain robustness.
Engineered a modular PyTorch pipeline with Hydra configs, Weights & Biases tracking, checkpointing, and a probabilistic head for uncertainty estimates.

TransformerPINNPyTorch

GitHub

02 · Full-stack

Mar 2026

ArthaSutra: Gamified Finance-Learning App

Next.js + Supabase · live on Vercel

Built a gamified finance-learning platform — virtual investing, league progression, and the Aarohan contest.
Engineered dual-mode persistence: Supabase auth + cloud sync, with a local-storage demo fallback when unconfigured.

Next.jsSupabaseRecharts

Live GitHub

03 · Research · Published

May 2024 - Present

EV Battery State-of-Charge Forecasting & Real-Time Inference

MAE 3.66 — beats prior benchmarks (4.28–6.5)

Developed a Bayesian-optimized LSTM model for EV battery SoC prediction using 70 real-world driving sessions.
Achieved MAE of 3.66, outperforming prior research benchmarks (4.28-6.5).
Engineered a real-time inference pipeline using pre-trained Keras models for continuous SoC monitoring.
Integrated SHAP-based explainability to analyze feature contributions in battery behavior.

LSTMBayesian OptimizationKeras

GitHub

04 · Recommender Systems

Mar 2026

CSAO: Cart-Aware Add-On Recommender

NDCG@10 0.707 · p95 latency <200ms

Designed a real-time food-delivery recommender — LightGBM LambdaRank retrieval, a neural cross-attention reranker, and LLM-generated explanations — served under 200ms.
Reached NDCG@10 of 0.707 (+15% vs popularity, +39% vs co-occurrence) with bootstrap 95% confidence intervals and Optuna hyperparameter tuning.
Validated cold-start (unseen user, restaurant, item) and catalog coverage on 952K synthetic order-items across 8 cuisines.

LightGBMLambdaRankLLM

GitHub

05 · Computer Vision · Production

Jun 2025 - Present

CAPTCHArd: CAPTCHA Recognition Service

OpenCV + TF · CI + signed model releases

Built a production Python library to segment and recognize 5-digit CAPTCHAs with OpenCV and TensorFlow, exposing both sync and async prediction APIs.
Hardened delivery with pytest and GitHub Actions CI, uv-locked environments, and SHA256-verified release-based model artifacts.

OpenCVTensorFlowCI/CD

GitHub

06 · Full-stack ML

StockVision: Real-Time Stock Forecasting Platform

LSTM · FastAPI + Streamlit + SHAP

Built end-to-end LSTM forecasting system using technical candlestick features.
Deployed Streamlit + FastAPI dashboard with SHAP-based model explainability.

LSTMFastAPIStreamlit

05 · research

Publication

published

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries

IEEE SEFET 2025 · IEEE Xplore

06 · recognition

Awards & certifications

IEEE Outstanding Student Volunteer Award for Academic & Extracurricular Excellence (June 2026)
IBM Data Science Professional Certificate (2025)
Lala Ram Kishan Das Jain Memorial Award for Initiative (2022)

07 · writing

Thoughts & perspectives

all posts

March 24, 2026

The Harness Problem: Why the Same Model Feels Like a Genius in One Tool and an Intern in Another

read article

March 16, 2026

Bias vs Variance: The Invisible Balance Behind Every ML Model

A clear and engaging explanation of bias, variance, underfitting, overfitting, and why building good machine learning models is really about finding the right balance.

read article

March 14, 2026

Build the MVP Before You Build the Dream

A powerful reflection on why the first version of a product matters more than the perfect version that never reaches users.

read article

March 10, 2026

A Holiday Accident That Quietly Broke Machine Learning's Oldest Rule

read article

contact

Let's build something intelligent.

Open to ML research collaborations and engineering roles. The fastest way to reach me is email.

aditya_mishra@outlook.in

Email LinkedIn GitHub

Machinelearningthatships.

Technical toolkit

Where I've worked

ML Engineer Intern

ML-IoT Intern

Life & leadership

Projects

Res-PhyT: Physics-Informed Transformer for Battery SoC

ArthaSutra: Gamified Finance-Learning App

EV Battery State-of-Charge Forecasting & Real-Time Inference

CSAO: Cart-Aware Add-On Recommender

CAPTCHArd: CAPTCHA Recognition Service

StockVision: Real-Time Stock Forecasting Platform

Publication

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries↗

Awards & certifications

Thoughts & perspectives

The Harness Problem: Why the Same Model Feels Like a Genius in One Tool and an Intern in Another

Bias vs Variance: The Invisible Balance Behind Every ML Model

Build the MVP Before You Build the Dream

A Holiday Accident That Quietly Broke Machine Learning's Oldest Rule

Let's build something intelligent.

Machinelearningthatships.

Technical toolkit

Where I've worked

ML Engineer Intern

ML-IoT Intern

Life & leadership

Projects

Res-PhyT: Physics-Informed Transformer for Battery SoC

ArthaSutra: Gamified Finance-Learning App

EV Battery State-of-Charge Forecasting & Real-Time Inference

CSAO: Cart-Aware Add-On Recommender

CAPTCHArd: CAPTCHA Recognition Service

StockVision: Real-Time Stock Forecasting Platform

Publication

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries↗

Awards & certifications

Thoughts & perspectives

The Harness Problem: Why the Same Model Feels Like a Genius in One Tool and an Intern in Another

Bias vs Variance: The Invisible Balance Behind Every ML Model

Build the MVP Before You Build the Dream

A Holiday Accident That Quietly Broke Machine Learning's Oldest Rule

Let's build something intelligent.

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries

A Bayesian-Tuned LSTM Approach to State of Charge Forecasting in Electric Vehicle Batteries