🚙 Soumyajyoti Dutta

Open to work · ML engineering & research roles

Hi, I'm Soumyajyoti.
I build ML systems that ship.

GitHub LinkedIn

CS PhD student at Texas A&M (Botacin's Lab) working on LLM post-training & alignment, evaluation design, agentic & RAG systems, and applied ML for cyberdefense. Models from 140M to 9B params, trained on SLURM clusters, shipped as tools people actually use.

✉️ Email me 📄 Resume 🎓 CSCE 313

Currently building

Targeting USENIX Security '27, Cycle 2

IOCScooper++ — IOC/TTP extraction from real threat reports

  • 50.6Mparams
  • 0.894micro F1
  • 36IOC types
  • 344rules generated
  • Trained ScooperNER, a 50.6M-parameter extractor (MLM pretraining → fine-tuning, 36 IOC types): 0.894 micro / 0.925 macro span-exact F1 on held-out vendor and government reports, beating fine-tuned encoder baselines at 3–7× its size, with 0.950 F1 on explicit MITRE ATT&CK IDs.
  • End to end on 47 held-out reports, extraction feeds frozen YaraGen at 0.86/0.88 precision/recall with 100% rule validity at 0.97 IOC coverage across 344 generated rules.
  • Extending extraction to grounded (entity, action, IOC) tuples lifted into ATT&CK techniques via procedure-example retrieval, open-set over all 637 techniques, with a full-report co-annotated benchmark of real malware analysis reports.

🗺️ My work, as a tiny road trip. Just keep scrolling; the car drives itself past seven stops. Hit 🌙 up top for a night drive.

Skip the drive, just read ↓

🧭 What I do

The short version: I make language models useful after pretraining, then measure whether they actually are. Most of that work runs on multi-node clusters, and most of it is aimed at security problems.

  • LLM post-training pipelines: curriculum/staged SFT, structural reinforcement, reward-signal design, and behavioral shaping for fine-tuning and alignment.
  • Evaluation frameworks: custom reward metrics (lexical / syntactic / semantic), LLM-as-judge and user simulation, human-curated benchmarking across 10+ LLM families, plus cross-modal vision-language eval.
  • Distributed & efficient training: 140M–9B-param families (incl. sparse-MoE) on multi-node SLURM clusters; DeepSpeed / FSDP / DDP-NCCL / CUDA; rank-stabilized LoRA; vLLM with multi-turn KV reuse.
  • Agentic & RAG systems: multi-tool dispatcher architectures, layered query routing, confidence-threshold anti-hallucination with web-search fallback, MCP-based tool integration.
  • Large-scale data pipelines: multi-source dataset construction (100M+ labeled examples), tokenizer design, checkpoint management, reproducible experiment tracking.
  • Systems & full-stack engineering: CUDA profiling, memory-efficient loaders, FastAPI/React apps, local LLM orchestration, ATS/scraper pipelines.

Stack: PyTorch · HuggingFace Transformers · TRL · DeepSpeed · FSDP · CUDA · SLURM · vLLM · Ollama · MCP · LightGBM/XGBoost · Python · C++ · Rust · FastAPI · React · SQLite · Playwright · YARA · malware analysis (PE features)

🏔️ Experience

2024 — present

Research Assistant / ML Engineer

Botacin's Lab, Texas A&M University

  • Built curriculum-based post-training pipelines across 140M–9B-param families (BART, T5, Gemma 3, LLaMA 3.x, Nemotron; incl. sparse-MoE) on multi-node SLURM clusters with DeepSpeed, FSDP, and CUDA.
  • YaraGen, a Booz Allen Hamilton sponsored, end-to-end LLM system for automated YARA threat-detection rule generation (USENIX Security 2027, submitted):
    • A 20-tool agent dispatcher with a 4-layer query router and a RAG pipeline with confidence-threshold anti-hallucination, web-search fallback, and a Nemotron-4B synthesizer.
    • A three-source dataset framework of 100M+ labeled examples, including 305K crawled real-world rules.
    • YaraAST, a hybrid AST-aware tokenizer (10× token reduction); post-trained models to ≥99% syntax / ≥95% logical validity on 22,655 held-out rules, a floor held even by BART-base 140M.
    • Scooper (BiLSTM hex extractor) and MultiRuleGen (IR-level merger) raised corpus coverage 83% → 100%.
    • Brownie & Puff, a reproducible reward/eval function (BLEU + parsing + SAT semantics) used as both training signal and quality gate; benchmarked 10+ LLM families over 150+ runs.
    • A custom NER pipeline for IOC extraction from analyst reports: 92.5% F1 across 25 IOC types.
  • AutoPYara, an open-sourced rule-generation package on PyPI (ACSAC 2026, accepted): re-implemented biclustering-based rule generation from 71K Windows PE samples; beat the AutoYara SOTA by +14% coverage, +10% threat-hunting accuracy, +8% low-sample generalization.

🌲 Projects

Built because I wanted them to exist. Most of them are still running.

Proactive Research Agent

Taught a deep-research agent to pause on underspecified queries and ask one clarifying question first. Trained Gemma-4-E4B (~9B sparse MoE) on a single 80GB GPU with a 4-stage SFT curriculum; lifted pause-detection F1 0.168 → 0.573 on a 176-query benchmark. Co-authored; led model development & training.

agentsSFT curriculumMCP tools2026
🏆 1st place

ML Malware Detection

Windows PE classifier over 2M+ samples (EMBER, BODMAS, Benign-NET): 96.2% accuracy, 0.97 F1 under 1GB/5s constraints, plus the competition's best-performing adversarial evasion suite. CSCE 689 ML-Based Cyberdefenses.

LightGBMXGBoostadversarial ML

Sturdy Fishstick

Self-hosted job & PhD search platform that scores listings with fully on-device local LLMs (no external inference API). Multi-source ingestion (ATS APIs, Google Jobs, LinkedIn), a categorical scoring scheme built for small models, and an FTS5-grounded RAG chat that never invents statistics.

FastAPIReactOllamaRAG

Multimodal & VL Evaluation

Evaluated a DeepSeek vision-language model against human-annotated ground truth; built the comparison harness and error analysis for a cross-modal eval exercise. ECEN 642, Digital Image Processing & Computer Vision.

VLMsevaluation

Multiperspective Hawkeye

The ISCA'16 Hawkeye cache-replacement policy implemented in zsim, extended with PC-tracking and multi-policy benchmarking under multi-threaded workloads.

zsimcomputer architecture

HelloPentagon

Explainable ML malware-defense chatbot: a classifier handles triage while an LLM interpretability layer explains the verdicts to analysts.

explainable MLsecurity

Also on the shelf

WildFire: spatiotemporal ML for wildfire-spread prediction from climate and satellite data · Carotid Artery USG Analysis: CNN ultrasound classification, published as a Springer book chapter (2022) · Fashion-MNIST: ResNet/VGG comparison with regularization analysis.

📄 Publications

  • YaraGen — USENIX Security 2027 (submitted).
  • AutoPYara — ACSAC 2026 (accepted); open-sourced on PyPI.
  • Training and Benchmarking Pro-active Deep Research Agents to Overcome Underspecificity in User Queries — manuscript in preparation, 2026 (co-authored, 4-person team).
  • Carotid-artery ultrasound CNN classification — Springer book chapter, 2022.

🎓 Teaching & outreach

TA — CSCE 313: Introduction to Computer Systems

Texas A&M · Fall 2026

Lab instructor and grader; I also write the small practical assignments that give students hands-on Linux experience. Course materials →

Learning outcomes
  • Understand what system software an application program needs to run, on a single node and over a network, and how that software is used.
  • Perform system-level programming.
  • Use system components: memory, file systems, process control, inter-process communication, and networking.
  • Understand the problems and pitfalls in designing and implementing multithreaded and networked applications.
  • Use services provided by the underlying operating system by programming directly at the OS interface level.
  • Understand these aspects of a computer system:
    • What an operating system is, and what its components are.
    • System calls and their purpose.
    • Execution of a program; function calls; interrupts.
    • The OS application interface, file system, and process control.
    • Inter-process communication.
    • Concurrency, threads, and process/thread synchronization.
    • Network socket programming.
    • Parallel and distributed computing.

TA — CSCE 439/704: Data Analytics for Cybersecurity

Texas A&M · Spring 2026

Designing assignments and mentoring student projects; clustering, supervised ML, anomaly detection, security-focused visualization.

Partner & Presenter — TAMU CS Day

2024 & 2025

Hands-on AI and cybersecurity demos for high-school students.

TA — ECEN 325: Electronics

Spring 2023

Lectures, grading, and lab sessions for analog electronics.

Technical Coordinator — Student Research Week

Summer 2023

Automated judging and submission infrastructure for one of TAMU's largest academic events.

Let's take a drive together 🚙

I'm open to ML engineering, research, and applied-science roles. If any landmark caught your eye, my inbox is the last stop on the map.

✉️ soumyajyotidutta23@gmail.com GitHub LinkedIn Resume (PDF)