Arnali Saha · Portfolio2026
charging
0%
initialising
M.Sc. Electrical Engineer · Karlsruhe Institute of Technology

ArnaliSaha

M.Sc. Electrical & Information Technology · KIT 2025

Electrical engineer working across battery diagnostics, electrochemical modelling, embedded systems, and physics-guided machine learning. My work focuses on extracting health, operational history, and future degradation behaviour from the electrochemical evidence a battery leaves behind.

Battery Electrochemical Modelling & EIS Hybrid Physics-AI Models Edge & Cloud Deployment
scroll
01 / The thread
Looks right. Is it?

Battery diagnostics appears straightforward. Measure a voltage, estimate a state, predict a future. In practice, every measurement is shaped by assumptions most people never examine.

Every measurement is influenced by chemistry, temperature, operating history, instrumentation, and modelling assumptions. Understanding a battery therefore requires more than prediction. It requires understanding the chain of physical processes that produced the measurement in the first place.

That means following one signal across the full vertical: the electrochemistry that produced it, the hardware that captured it, the firmware that processes it, and the models, both physics-based and machine-learned, that try to explain it. That range isn't scatter. It's the depth a single noisy measurement actually demands.

Disagreement between measurement and prediction is rarely random. It usually points toward missing assumptions, incomplete models, or physical effects that have not yet been accounted for.

01

Models are only useful where they fail predictably.

02

Every measurement carries assumptions. Find them before you trust it.

03

Uncertainty should be quantified, not concealed.

02 / The stack

I work the whole
vertical of one problem.

From the molecules inside a cell to a diagnosis running in the cloud. These are layers of the same system, not separate interests.

01

Electrochemistry & Diagnostics

EIS measurement & spectral analysis · ECM parameterisation · SEI / charge-transfer / diffusion signatures · SOH, SOC & RUL estimation.Linking measurable signatures to underlying electrochemical mechanisms.
02

Embedded & Hardware

HV-BMS PCB design · protection circuitry · embedded C++ · ESP32 + ADS1115 pulse-test instrumentation · FPGA exposure.Measurement quality bounds everything downstream.
03

BMS & Validation

HiL / SiL test environments · ISO 26262 & ISO 12405 · CAN / Vector CANoe · onboard vs offboard model comparison.Characterising behaviour under real operating conditions.
04

Data & Hybrid Physics-AI

Python pipelines · physics-informed ML (gradient boosting, RF, SVR) · anomaly detection · PyBaMM · LLM-assisted abductive inference.Combining mechanistic understanding with statistical generalisation.
05

Edge & Cloud Deployment

Docker · CI/CD · real-time algorithms on constrained hardware · digital-twin prototypes · API integration.From research prototype to deployable diagnostic system.
03 / The journey

Let's start at the
beginning.

Every role I took was an answer to a question the last one couldn't close.

2012–2016

How does a real circuit betray its own schematic?

B.Tech Electronics & Instrumentation · Techno India College, Kolkata

Hardware never quite behaves the way the textbook promises, and that gap is where the learning lives. Built robotics, chased the difference between theory and bench, published a first paper before graduating.

EmbeddedRoboticsPublication
2016–2018

Reliability is rarely determined by a single design decision. It emerges from thousands of small ones.

Electrical Design Engineer · Powertech Engineers, India

Schematics, control diagrams, industrial automation, all verified against safety standards. The difference between "functional" and "reliable" became something close to an obsession.

Electrical DesignIndustrial Automation
2019–2025

A degree that kept dismantling what I thought I already knew.

M.Sc. Electrical & Information Technology · KIT, Karlsruhe

Every semester raised a question the previous one couldn't answer. It ended in a machine-learning thesis validated inside a real automotive production environment, the framework underneath everything else on this page.

the spine of the story
2019–2020

Race cars are remarkably efficient at exposing bad engineering.

HV-BMS Design · KA-RaceIng (Formula Student), KIT

Designed the high-voltage BMS for an electric race car from scratch: PCB hardware, protection circuitry, embedded C++ firmware, CAN integration, on-track commissioning. Shipped something that had to be right under load, not in a slide.

components → system behaviour
HV-BMSPCBEmbedded C++
2020–2021

The challenge was no longer measurement. The challenge was determining which measurements could be trusted.

Research Assistant · Energy Storage Systems (IAM-ESS), KIT

First sustained exposure to real electrochemical data: noisy, ambiguous, and refusing to match the model cleanly. Built Python pipelines to chase the disagreement between model and reality.

where measurement & model first disagreed
2022–2023

Does the battery actually do what the spec promises?

e-Drive & BMS Validation · Mercedes-AMG, Affalterbach

System-level HiL / SiL validation of safety-critical BMS and e-drive systems. An exploratory thread, estimating thermal state from impedance spectroscopy with no extra sensor, taught me more about electrochemical measurement than any course.

validation → sharper questions
HiL / SiLISO 26262CANoe
2023–2024

Impedance spectroscopy became a language rather than a measurement technique.

Battery Algorithms · Mercedes-Benz · Electrochemical Technologies (IAM-ET), KIT

Built ECM-based power and resistance prediction directly from impedance spectra, then ran EIS and polarisation campaigns across batteries and PEM fuel cells. The chemistry changed; the core analytical question did not.

impedance as a language
EISECMP2D calibration
2024–2025

Where do physics-based models break, and what does that boundary tell us?

Master's Thesis · Battery SOH, Degradation & RUL · Mercedes-Benz

A hybrid physics-and-ML framework for SOH, degradation-trajectory and RUL estimation across large automotive lifecycle data. The key finding: cells with identical cumulative usage aged differently depending on their history. Path matters. >98% model-measurement correlation, <4% error, inside an ISO 26262 confidence framework.

the answer to everything since IAM-ESS
Hybrid Physics-AIPath-Dependent AgeingRUL
2026 →

If a cell arrives with no history at all, can we reconstruct the life it lived?

Retrograde · Independent Research

The thesis proved that path determines a cell's future degradation trajectory. Retrograde asks the inverse, and harder, question: with no documentation and no baseline, can you reason backwards from the cell's present electrochemical state to the operational history that produced it?

04 / Flagship · Independent research · 2026

Retrograde.

Electrochemical forensics for batteries with no past.

Domain Second-life battery assessment
Method Physics-constrained abductive inference
Model Equivalent Circuit + Large Language Model
Hardware Custom HPPC platform · salvaged NMC pouch

The battery may not have a logbook.
That doesn't mean it has no history.

Two cells. Identical State of Health. Completely different futures.

Two batteries can exhibit identical state-of-health estimates while possessing fundamentally different degradation trajectories. Capacity alone does not reveal how a battery arrived at its current condition.

Fast charging, thermal stress, storage behaviour, and cycling intensity may all leave distinct electrochemical signatures despite producing similar present-day health estimates. Every existing diagnostic assumes a known, continuous history. Salvaged cells don't have one.

Cell A · gentle calendar ageing80%
Cell B · aggressive fast-charge80%
Same SOH · not the same future

The idea: reason backwards
from the evidence.

A salvaged cell isn't a system with missing data; it's a physical record of prior operating conditions. Every stress event left a trace. Some signatures emerge from SEI growth, others from lithium inventory loss, impedance growth, or prolonged thermal exposure. Retrograde explores whether parts of that history can be reconstructed from present-day electrochemical measurements alone.

Problem

State of health tells us where a battery is. It says nothing about how it got there, and how it got there determines its future degradation trajectory.

Approach

Extract electrochemical fingerprints via HPPC and incremental capacity analysis. Run abductive inference over a physics-scored hypothesis space to reconstruct probable operational history.

Outcome

Differentiate batteries that appear identical today but will age differently tomorrow. Confidence-bounded output, so the system knows when it doesn't know.

Architecture
Battery
HPPC + ICApulse characterisation
ECM ExtractionR₀ · R₁C₁ · diffusion terms
Digital Fingerprintelectrochemical signature
Hypothesis Generatorcandidate operational histories
Physics-Constrained Validationresidual · rank · reject
History Reconstructionconfidence-bounded output

Live diagnostics

A cell in active observation. Pulse test, ECM extraction, physics scoring: every change in the residual is a question.

retrograde_log · cell_18650_S01.out
Physics scoring
Physics match score···select
Measured cell Predicted by hypothesis

The result I'm proudest of isn't a clean answer.

Validating against cells with known ground-truth histories, Retrograde correctly recovered cycling, calendar, and shallow-use archetypes. But one cell, CS2_34, was genuinely ambiguous, and the system returned low confidence across every hypothesis rather than forcing a tidy story onto it. Outputs are framed as degradation phases, not precise dates, to keep every claim defensible.

// built with

PythonNumPySciPyPyBaMMStreamlitDockerLLMESP32 + ADS1115NetworkX
05 / Beyond the bench

The discipline doesn't
switch off at the lab door.

Each of these teaches the same temperament I bring to engineering: precision, structure, and a refusal to settle for "close enough."

Olympic Air-Pistol Shooting

Kreismeisterin 2026

Competing at the Landesmeisterschaft Baden-Württemberg. Olympic precision shooting is its own kind of systems engineering: the nervous system, the breath, the trigger finger, all tuned to fire within a tolerance most instruments can't measure.

Ten metres, one shot, zero room for a hopeful guess.

Music

Vocalist · Bassist · Songwriter

Writing and performing in indie / alternative. Composition is pattern recognition under constraints, the same instinct that reads a noisy impedance spectrum and hears something useful in it.

Structure you can feel before you can explain it.

Design & 3D

Graphic Design & 3D Modelling

Visual communication of complex systems, including this site. Making something technically dense immediately legible is one of the hardest problems I know.

Making the complicated legible is its own kind of engineering.

Build for the joy of it

IoT · AI Agents · Automation

ESP32 sensor systems, LLM agent integration, and workflow automation pipelines. The constraints of a weekend build are where the most honest engineering decisions get made.

Weekends are where I test the questions the week couldn't answer.

06 / Now

Looking for problems
that aren't fully
defined yet.

  • Battery diagnostics and prognostics
  • Physics-informed machine learning
  • Battery history reconstruction
  • Embedded diagnostic instrumentation
  • Second-life battery intelligence

Where I fit

BMS & state estimationBattery diagnostics & researchE/E validation & integrationEdge & connected battery intelligenceCloud battery platforms

Location Based in Karlsruhe, Germany.
Relocation Open to anywhere in Germany and across Europe.

07 / Contact

Let's talk.

Portrait of Arnali Saha
Arnali SahaKarlsruhe, DE · 2026
For roles, conversations, collaboration

I am particularly interested in battery diagnostics, battery systems, validation, embedded engineering, and technologies at the intersection of physical systems and intelligent software. If you are working on difficult problems in any of those domains, I would be glad to hear from you.

Write to mearnalisaha@gmail.com