INDEX / 2026DATA SCIENCE · NLP · POST-TRAININGBENGALURU 12.97°N
Soumyajit Roy.
Models that make the call. Research on how LLMs fail.
I'm Soumyajit, a data scientist at Pine Labs in Bengaluru. I work across the full arc of a model, from problem framing and data pipelines to deployment and monitoring, in e-commerce risk, credit underwriting and payments infrastructure. Alongside that, I research how language models fail and what post-training can do about it.
CURRENTLYDATA SCIENTIST I AT PINE LABS · SINCE AUG 2025
for people
Hi, I'm Soumyajit.
At work I decide things: whether an order can ship cash-on-delivery, what someone earns according to a bank statement nobody formatted, how much infrastructure payment traffic will need tomorrow.
In research I look at where language models go wrong, and what LoRA, loss design and ensembling can do about it without needing a bigger model.
BengaluruACL · AACL-IJCNLPB.TECH CSE '25
IN PRODUCTION AT
[ 01 / APPLIED ]
All work →In production
A.01
Checkout risk
Return-to-origin models that decide, at checkout, whether an order can go cash-on-delivery — including for shoppers and merchants we've never seen before.
A.02
Lending signals
NLP on messy UPI bank statements for SignalIQ: finding salaries, recurring obligations and income sources that older parsers miss.
A.03
Payments infra
Traffic forecasting that keeps infrastructure provisioning honest — enough headroom, no waste.
[ 02 / RESEARCH ]
All research →Selected papers
[ TIMELINE ]
SEP 2026Five years, scrubbable
Drag the playhead to move through roles, papers and milestones.
2021
2022
2023
2024
2025
2026
PAPERJUL 2026
ModusPonens & Ambirig at SemEval-2026
Two system papers at ACL: belief bias in LLMs, and ordinal word-sense modelling.
ROLEData Scientist I · Pine Labs
PAPERS5
WORKPAPEREDUCATION
[ 03 / STACK ]
The stack, mapped to the work
Each tool is shown against the projects it was used on. Hover over a row or column to trace it.
HOVER THE GRID
PINE LABS · PRODUCTION
RESEARCH
TOOL ↓ · PROJECT →
Σ
PythonLANG
9
CatBoostGBM
2
XGBoostGBM
1
Logistic regressionML
1
BERTNLP
1
DeBERTa-v3NLP
1
Qwen-2.5-14BLLM
1
Gemma-2-9BLLM
1
LoRA / QLoRAFINE-TUNE
2
Vision-language modelsVLM
1
Model ensemblingML
1
ALSOSQLPANDASNUMPYLIGHTGBMPROMPT ENGINEERINGAWS SAGEMAKERAWS S3GIT / GITHUB / GITLAB