Shubhankar
Tiwari
Building intelligent ML systems that observe, detect, and adapt in real time.
I focus on:
- [01]Runtime Intelligence — controlling and stabilizing LLM behavior during generation
- [02]Model Lifecycle Intelligence — detecting drift and governing ML models in production
Project Spotlight
Built a full-stack system to monitor and control LLM generation at the token level, with adaptive instability detection and real-time intervention
LLM Control System — Runtime Reliability for Generative Models
LLM Control System — Runtime Reliability for Generative Models
- •Token-level observability via manual decoding loop with per-token probability and entropy extraction
- •Three-category stability detector system: entropy_collapse, repetition_loop, low_entropy_lock
- •Formal policy controller with explicit intervention rules mapped to instability signals
- •Adaptive regeneration and temperature adjustment — mid-generation intervention without restart
Core Insight
Detect and correct unstable token generation in real time using entropy-based control.
How I Think About ML Systems
Most ML systems fail silently.
I design systems that:
- measure behavior (entropy, drift, performance)
- detect failure modes (instability, degradation)
- take corrective action automatically
This results in ML systems that are not just predictive — but self-aware and controllable.
My Journey
I went from curious undergrad to backend engineer at Bank of America. The path wasn't linear — four years of production systems, a Kaggle Expert rank, and now building evaluation infrastructure for LLMs. Each of those things informed the others more than I expected.
B.Tech CSE
SRM Institute of Science & Technology
Graduated with 94.4% GPA in Computer Science. Built foundation in algorithms, data structures, and software engineering.
Software Engineer
Bank of America
Building and maintaining Java backend systems for corporate banking. I own the full lifecycle — design, CI/CD, deployments, and being on-call when something breaks. Four years of that teaches you things about software that writing code in isolation doesn't.
Kaggle Expert
Top 4.1% Globally
Notebooks Expert ranked #2,441 / 59,663 — personal best #707. 34 notebooks and 10 bronze medals across ML, DL, NLP, Computer Vision, and regression.
Backend Excellence
Currently at Bank of America, I focus on building microservices that power critical banking workflows. From API design to deployment validation, I own the full lifecycle.
Competitive ML
Outside of work, I sharpen my skills through Kaggle — ranked top 4.1% globally, personal best #707. Currently researching LLM behavioral reliability, cultural alignment, and building production safeguard systems that measure deployment-readiness, not just benchmark accuracy.
Engineering at Scale
At Bank of America I work on backend systems that process real financial transactions. My job isn't just writing code — it's release validation, production triage, and being accountable when a deployment goes wrong at an inconvenient hour.
I've been doing this for four years. The thing that changes is your relationship with failure. You stop being surprised by it and start building systems that tell you clearly when and why they're failing.
Engineering Principles
Projects & Experiments
Selected projects and experiments
Built an end-to-end ML system that detects data drift, evaluates model degradation, and governs retraining decisions.
Core Insight
Link distributional drift to model degradation and automatically recover performance via retraining.
Production validation system for AI-generated outputs in financial services — threshold optimisation, conservative AND escalation policy, priority-scored human review queue, and audit-grade provenance.
End-to-end multilingual LLM training pipeline targeting Hindi/English code-switching. Dataset curation, LoRA fine-tuning, inference evaluation, and deployment packaging across 6 Kaggle notebooks.
ML-based workout song recommender using BPM and VADER sentiment analysis. Co-authored research with K-Means clustering on Billboard Top 100 to match songs to exercise intensity.
Notebooks Expert rank #2,441 / 59,663 — personal best #707. 34 notebooks, 11 datasets, 3 models, 1 competition entry. 10 bronze medals across ML, DL, NLP, Computer Vision, and regression.
System-Level Focus
These projects represent three pillars of a coherent ML systems narrative, with decision-aware control and auditability:
- [01]Inference: LLM Control System → Runtime behavioral guardrails
- [02]Evaluation: Fraud System → Drift detection and automated lifecycle management
- [03]Governance: LLM Output Validation System → Post-deployment validation layer for LLM outputs
Together, they reflect a focus on building ML systems that are:
observable, controllable, and production-ready.
Technical Focus
ML Systems
Modeling
Infrastructure
Credentials
Kaggle Notebooks Expert
Rank 2,441 / 59,663 · 10 Bronze Medals
B.Tech CSE — SRM IST
94.4% GPA · 2018–2022
ML, Deep Learning, CV
Kaggle Certified · 2025
Interactive Terminal
Try: help, about, ls projects/, cat skills, neofetch
Get in touch
Open to roles in backend systems, platform engineering, and applied AI. If something here resonates — I'd like to hear from you.
tiwarishubhankar@gmail.com