Prateek S. profile photo 1 on Tal

Prateek S.

AI Architect | Enterprise ML Systems | Driving Strategy Through Data | IIT Kharagpur | US Patents

Greater Bengaluru Area

10+ years of work experience

Accenture
Mar 20267 moPresent

Bangalore

Published CHIAR-Former (arXiv:2606.08327): novel ~400M-parameter efficient transformer routing tokens between DCT spectral mixing and full self-attention via per-token spectral entropy; WikiText-103 Test PPL 23.58 at 37% FLOP reduction. Published LayerRoute (arXiv:2606.01838): frozen Qwen2.5-0.5B backbone with per-layer hard-gated skip connections; 12.91% FLOP skip differential (tool calls 15.25% vs planning 2.34%), trained in 6.4 min on A100 with 0.22% trainable params. Conducting Target-Space Distillation (TSD) experiments on TinyLlama-1.1B comparing TSL vs SFT students; TSL Val PPL 5.44–7.68 vs SFT 5.99–10.09 (SFT overfits from epoch 2).
HP Inc. (R&D Center)
May 20196 yrs 9 mosJan 2026

Bengaluru, Karnataka, India

Built Transformer-based models for time-series anomaly detection and throughput prediction across printer manufacturing telemetry — boosted system throughput 30% via learned sequence representations. Led VAE (Variational Autoencoder) anomaly detection research across enterprise manufacturing datasets, achieving defect rate under 2% — directly applicable to ad-quality and fraud detection pipelines. Deployed adaptive A/B testing framework using contextual multi-armed bandit algorithms (ε-greedy, Thompson Sampling) to dynamically route inference across 15+ model variants, improving accuracy 8% — analogous to ad-serving model selection. Fine-tuned BERT/RoBERTa/DistilBERT for multi-class document relevancy classification and content scoring, achieving 94% accuracy (F1-optimised) on 24K+ image/document datasets across 15 web categories — directly transferable to ad content classification and brand-safety scoring. Built XGBoost classification pipelines with chronological splits and SMOTE for imbalanced data; strong out-of-time performance on tabular prediction tasks comparable to CTR/CVR modelling requirements. Architected Agentic RAG system using LangGraph and multi-agent orchestration for printer support automation, reducing customer response time 60%; evaluated via LangSmith on golden query dataset. Built end-to-end MLOps pipelines (MLflow, Docker) covering model deployment, real-time monitoring, and champion-challenger A/B frameworks; contributed to 3 U.S. patent filings in deep learning for predictive maintenance.
Citi
Aug 20171 yr 10 mosMay 2019

Bangalore

Citi
Aug 20171 yr 9 mosApr 2019

Built deep learning targeting models (TensorFlow, PySpark) for credit cards; achieved 22% response rate improvement through customer segmentation and propensity modelling. Built XGBoost response models saving $1M+ in targeting costs through automated customer selection.
Equifax
Jun 20161 yr 2 mosJul 2017

Managed credit risk scorecards (logistic regression, gradient boosting) improving KS by 12 points; built e-commerce seller scorecards (Random Forest, 72% accuracy).
Equifax India
Jun 20161 yr 2 mosJul 2017

Machine Learning projects
GitHub
@hackie123
Medium
@prateeksikdar

Prateek’s AI Stack

claude code
claude code
hugging face
hugging face
langgraph
langgraph
Qwen
Qwen

Education

Indian Institute of Technology, Kharagpur

B.Tech (Honors)

Indian Institute of Technology, Kharagpur

20044 yrs2008

I.C.S.E.

St. Vincent's Convent School

Join Prateek and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Prateek S. on Tal

Headline

AI Architect | Enterprise ML Systems | Driving Strategy Through Data | IIT Kharagpur | US Patents

About

AI Architect with 14+ years of experience building scalable machine learning systems across banking and enterprise technology. My work spans experimentation at scale, intelligent systems, and AI-driven optimization. I am deeply interested in how strategy, data, and emerging technologies reshape competitive advantage — and how leaders can harness them responsibly for long-term value creation.

Location

Greater Bengaluru Area

AI stack

claude code

hugging face

langgraph

Qwen

Work experience

AI Architect and Researcher at Accenture, Mar 2026 - Present · 7 mo, Bangalore, Published CHIAR-Former (arXiv:2606.08327): novel ~400M-parameter efficient transformer routing tokens between DCT spectral mixing and full self-attention via per-token spectral entropy; WikiText-103 Test PPL 23.58 at 37% FLOP reduction. Published LayerRoute (arXiv:2606.01838): frozen Qwen2.5-0.5B backbone with per-layer hard-gated skip connections; 12.91% FLOP skip differential (tool calls 15.25% vs planning 2.34%), trained in 6.4 min on A100 with 0.22% trainable params. Conducting Target-Space Distillation (TSD) experiments on TinyLlama-1.1B comparing TSL vs SFT students; TSL Val PPL 5.44–7.68 vs SFT 5.99–10.09 (SFT overfits from epoch 2).

Machine Learning Researcher at HP Inc. (R&D Center), May 2019 - Jan 2026 · 6 yrs 9 mos, Bengaluru, Karnataka, India, Built Transformer-based models for time-series anomaly detection and throughput prediction across printer manufacturing telemetry — boosted system throughput 30% via learned sequence representations. Led VAE (Variational Autoencoder) anomaly detection research across enterprise manufacturing datasets, achieving defect rate under 2% — directly applicable to ad-quality and fraud detection pipelines. Deployed adaptive A/B testing framework using contextual multi-armed bandit algorithms (ε-greedy, Thompson Sampling) to dynamically route inference across 15+ model variants, improving accuracy 8% — analogous to ad-serving model selection. Fine-tuned BERT/RoBERTa/DistilBERT for multi-class document relevancy classification and content scoring, achieving 94% accuracy (F1-optimised) on 24K+ image/document datasets across 15 web categories — directly transferable to ad content classification and brand-safety scoring. Built XGBoost classification pipelines with chronological splits and SMOTE for imbalanced data; strong out-of-time performance on tabular prediction tasks comparable to CTR/CVR modelling requirements. Architected Agentic RAG system using LangGraph and multi-agent orchestration for printer support automation, reducing customer response time 60%; evaluated via LangSmith on golden query dataset. Built end-to-end MLOps pipelines (MLflow, Docker) covering model deployment, real-time monitoring, and champion-challenger A/B frameworks; contributed to 3 U.S. patent filings in deep learning for predictive maintenance.

Data Scientist at Citi, Aug 2017 - May 2019 · 1 yr 10 mos, Bangalore

Data Science Manager at Citi, Aug 2017 - Apr 2019 · 1 yr 9 mos, Built deep learning targeting models (TensorFlow, PySpark) for credit cards; achieved 22% response rate improvement through customer segmentation and propensity modelling. Built XGBoost response models saving $1M+ in targeting costs through automated customer selection.

Data Science Manager at Equifax, Jun 2016 - Jul 2017 · 1 yr 2 mos, Managed credit risk scorecards (logistic regression, gradient boosting) improving KS by 12 points; built e-commerce seller scorecards (Random Forest, 72% accuracy).

Data Scientist at Equifax India, Jun 2016 - Jul 2017 · 1 yr 2 mos, Machine Learning projects

Education

B.Tech (Honors), Indian Institute of Technology, Kharagpur

I.C.S.E., St. Vincent's Convent School

Public links

github: https://github.com/hackie123

medium: https://medium.com/@prateeksikdar