AC Loading

ANGSHUMAN
CHAKRAVERTTY

End-to-end pipelines, benchmarking frameworks,
and intelligent agents.

View My Work Download CV

Driven by curiosity.
Powered by data.

I'm an ML Systems and Data Science Engineer building end-to-end pipelines at the intersection of scientific computing, generative AI, and MLOps. My work ranges from fine-tuning genomic language models on ancient DNA to benchmarking LLMs on hardware design tasks — two lines of work now written up as arXiv preprints — always with an eye on reproducibility and real-world deployment.

Most recently I shipped backend AI services in production — multi-language OCR, async batch pipelines, and document Q&A — where correctness under real rate limits and messy Unicode mattered more than benchmark numbers.

I care about the full lifecycle: research, model design, containerised deployment, and production monitoring — a Kubernetes operator that survives a real EC2 Spot interruption is the same discipline as a retrieval model that survives real telescope data. If a system doesn't hold up outside a notebook, it isn't done.

Genomic ML LLM Systems MLOps Benchmarking Sim-to-Real Scientific Computing
Angshuman Chakravertty
0 + GitHub Repos
0 + Commits
0 + Technologies
0 arXiv Preprints

Research & Preprints

arXiv:2607.27712 [cs.LG] 2026

VESTIGE: A Knowledge-Guided Masking Strategy for Corruption-Aware Fine-Tuning of Genomic Transformers, Validated on Ancient DNA Reconstruction

Chakravertty, A., Maheshwari, R.

  • Introduced a parameter-free masking strategy that aligns training with an empirically measured per-position corruption profile, replacing the standard assumption that reconstruction difficulty is position-agnostic.
  • Cut validation cross-entropy 13% (3.274 vs. 3.757) and led standard MLM at every damage width tested (Δ = +4.18 to +10.35 pp, all p < 10⁻⁸), with ESMFold reconstructions holding TM-score > 0.95 even under damage amplified up to 30× beyond authentic rates.
arXiv → Case study → GitHub →
arXiv:2607.22759 [cs.AR] 2026

Benchmarking LLMs for Verilog Design Flows

Chakravertty, A., Koshti, R., Sharma, B.P., Chamola, V.

  • Built a reproducible benchmarking platform evaluating open-source LLMs on Verilog RTL generation across 50 curated tasks spanning combinational, sequential, FSM, and mixed designs — validated via Verilator compilation and Icarus simulation.
  • Raised syntax validity from 0% to 70.4% average and simulation pass rate to 51.8% across 1,610 runs on three open-source LLMs (Llama-3-8B, StarCoder2-7B, TinyLlama-1.1B); prepared for submission to IEEE Design & Test.
arXiv → Case study → GitHub →

In Production

Jun — Jul 2026 Hyderabad, IN

AI Backend Engineering Intern

Asset Telematics Pvt Ltd Certificate GitHub

  • Extended a Gemini-powered OCR microservice to emit structured output in five languages (English, Arabic, Hindi, French, German) across 10 document types, via prompt engineering and a per-request output_lang parameter. Traced a silent Unicode data-loss bug to a Latin-only regex that was stripping Arabic and Devanagari — replaced it with Unicode-category-aware cleaning that preserves combining marks.
  • Built an async batch-processing pipeline (/extract_batch + /batch_status polling) with bounded parallelism via an asyncio semaphore and per-file partial-failure isolation — engineering concurrency around a blocking vendor SDK using a multi-key round-robin thread pool, validated to degrade gracefully under live API rate limits.
  • Integrated OpenRouter (Qwen3-VL) as an isolated document Q&A module with upload-once/ask-many sessions, plus a benchmarking harness measuring throughput, latency, and extraction accuracy — ~86% name-extraction accuracy on passports, with structured validation rejecting incomplete documents.
Python FastAPI asyncio Gemini OpenRouter Qwen3-VL

Selected Work

01 2026

ANVIL

Autonomous ML red-teaming — an 8-phase pipeline that attacks any PyTorch model with 4 attacks written from scratch, clusters failure modes with UMAP + HDBSCAN, explains them with a RAG-grounded LangGraph agent over 10 adversarial ML papers, patches behind a safety gate (score ≥ 0.70, accuracy drop ≤ 3%), and ships a PDF audit.

PyTorch LangGraph FAISS UMAP HDBSCAN FastAPI
02 2026

ARGUS

Spot-resilient ML training orchestrator — a Kubernetes operator (SpotResilientJob CRD + kopf reconcile loop) that survived a real EC2 Spot interruption on EKS: drained in 10s, checkpointed a live PyTorch job to S3 on SIGTERM, and resumed on a healthy node in ~75s with one epoch of work lost.

Kubernetes kopf AWS EKS Terraform Prometheus Grafana
03 2026

MIRAGE

Sim-to-real exoplanet atmosphere retrieval — a multi-instrument transformer encoder with a noise-conditioned flow-matching posterior. Diagnosing the WASP-39b collapse as a radius/baseline degeneracy drove best-fit χ²/dof from 301 → 0.06 and posterior coverage 2/7 → 7/7, then generalised unchanged across 3 targets.

PyTorch Flow Matching TauREx3 Optimal Transport JWST

Tech Stack

Languages

Python C++ SQL

ML / DL

PyTorch TensorFlow Scikit-learn HuggingFace Transformers

Gen AI & LLMs

LangChain LangGraph RAG Pipelines Ollama

Data & Viz

Pandas NumPy Matplotlib

MLOps / DevOps

Docker Kubernetes Terraform Prometheus Grafana Git & GitHub CI/CD (AWS CodePipeline) Linux/Bash

Cloud & Deployment

AWS EC2 S3 EKS SageMaker Elastic Beanstalk FastAPI Flask Streamlit

Academic Background

B.Tech, Computer Science & Engineering (Data Science)

SVKM's NMIMS — Hyderabad

Jul 2023 — May 2027 CGPA 7.1 / 10 · Sem 5 GPA 3.14 / 4.0

Achievements & Certifications

2025

1st Place — NMIMS Hackathon NMTF

Team of 6 · 24-hour sprint

2026

Open Source Contributor — ML4SCI / DeepLense

PR #194 torchmetrics API fix · PR #195 dynamic --entity CLI arg · PR #190 dataset setup docs

2024

CS50's Introduction to AI with Python

Harvard / edX · GitHub repo →

2024

Python for Data Science

NPTEL · IIT Madras

2024

Introduction to Machine Learning

NPTEL · IIT Kharagpur

Let's Build
Something Together

Open to collaborations, research opportunities, and interesting problems.

Copied to clipboard ✓