open to applied scientist / ml engineer roles · nyc / remote

Nirman Taterh

I build language systems — and the evaluations that keep them honest.

ML engineer across NLP, RAG, and agents. Transformer interpretability research on one side, deployed production systems on the other — the evaluation harness always comes first. M.S. Data Science, NYU.

+18%model accuracy@ J&J
−20%LLM hallucinations@ Miko
+50%click-through rate@ Last.fm
Download resumeEmail me

ai assistant trained on my work — ask it anything ↘

01 — work

Selected Work

NLP Research · featured

Prosodic Encoding in LLMs

Do transformers learn prosody without ever hearing speech? Linear probes across all BERT layers on the Helsinki Prosody Corpus, with POS-residualized controls to isolate prosodic signal from syntax. ACL-tier submission in progress.

GitHub ↗
PyTorchTransformersProbingInterpretability
QueryKeythesoundofhervoicerose
attention map
GoalPlannerExecutorCriticControllerstateadaptive loop · 88 tests · 8/8 benchmark
agent loop

Agents

Orchid — Multi-Agent Framework

Orchestration framework that adapts strategy mid-run: confidence tracking, critic debate panels, dynamic specialist delegation. Controller is itself an agent. No LangChain; runs fully local via Ollama. 88 tests.

GitHub ↗
PythonOllamaMulti-AgentPydantic
LangGraph orchestrationQueryBGE-M3hybridQdrantRRFColBERTrerankLLMgenerateMLflow · faithfulness gate · FastAPI
RAG pipeline

RAG / Systems

Production RAG Pipeline

Agentic RAG with LangGraph, BGE-M3 hybrid retrieval, Qdrant, and ColBERT reranking. 0.92 Recall@10 on BEIR SciFact — hybrid fusion +4 pts over dense-only, rerank +1.4 MRR. Faithfulness gate with automatic retry; MLflow tracking.

GitHub ↗
LangGraphBGE-M3ColBERTQdrant
modelfraud ◆legitimate
fraud detection network

Systems

Large-Scale Review Trust Modeling

Fake-review detection across 26.7M Amazon reviews: 97.7% recall / 96.1% accuracy vs rule-based baselines (logistic regression, random forests), then BiLSTM + attention with metadata fusion reaching 0.966 ROC-AUC. Fine-tuned DeBERTa cut sentiment-scoring MAE from 1.47 to 0.46 (−69%) via trust-weighted composite scoring. Leak-safe evaluation protocol and calibration auditing reported alongside headline numbers.

GitHub ↗
TensorFlowBiLSTM + Attentionscikit-learn
time steps →B/HRL
performance vs baseline

Published Paper

FinRL Crypto Trading Agent

Springer-published RL trading agent with FinBERT sentiment features. Beat benchmark strategies by 10%.

arXiv ↗Springer ↗
FinRL-XFinBERTPublished

02 — research

Research Focus

LLM Trustworthiness

Eval

Benchmarking reliability, calibration, and faithfulness of language models

Agent Evaluation

Eval

Evaluation infrastructure for reasoning models and agents

Agentic Orchestration

Agents

Autonomous planning, reflection, and mid-run strategy adaptation

Mechanistic Interpretability

Research

What linguistic structure transformers encode, and where

Memory & Retrieval

Systems

Production agents combining memory, retrieval, and tool use

03 — experience

Work History

Johnson & Johnson

Mar – Aug 2025

New York, NY (Remote)

Data Science Intern — LLM/NLP

  • ·Improved biomedical decision-support model accuracy by 18% by identifying evaluation gaps and iterating transformer-based NLP pipelines end-to-end
  • ·Cut preprocessing latency from 45s to 31s (−30%) by proactively redesigning the ETL stack before bottlenecks were flagged, improving reliability across 3 biomedical R&D workflows
  • ·Accelerated team experimentation by 25% by introducing AI-assisted coding workflows and building shared evaluation tooling adopted org-wide

Miko

Nov 2024 – Jan 2025

California, USA (Remote)

Conversational AI Developer

  • ·Reduced LLM hallucinations by 20% across 5,000+ robot interactions by independently redesigning RAG pipelines and tool-calling interfaces
  • ·Improved system throughput by 25% and cut error rate by 15% by building modular Python evaluation pipelines and automating model optimization loops

Miko

Jun – Sep 2022

Mumbai, India

NLP Intern

  • ·Improved multilingual classification accuracy by 25% and response quality by 35% by prototyping transformer pipelines, owning fine-tuning end-to-end

Paramount Global (Last.fm)

Sep 2020 – Feb 2021

London, UK (Remote)

Music Recommendation Systems Engineer

  • ·Increased click-through rate from 2.8% to 4.2% (+50%, p < 0.05) across 1.9M users with a hybrid recommender (collaborative filtering + NLP metadata features) for 93K artists
  • ·Reduced inference latency by ~30% by deploying distributed training across 4 GPUs and redesigning the data pipeline

04 — capabilities

Technical Skills

LLM Systems

Fine-tuning, evaluation, and deployment of transformer-based models.

PyTorchHuggingFaceDeBERTa-v3LoRA/PEFT

Retrieval & Agents

RAG pipelines and agent orchestration for production retrieval workloads.

BGE-M3ColBERTQdrantRAGatouilleLangGraph

Research & Evaluation

Benchmark design, error analysis, and evaluation harnesses.

ProbingEval designError analysisMLflow

Infrastructure

Scalable ML systems, inference APIs, and experiment tracking.

FastAPIDockerAWSSparkKubernetes

Reinforcement Learning

Policy optimization for sequential decision making and recommenders.

FinRL-XSACTD3Stable Baselines 3

05 — contact

Get In Touch

Looking for Applied Scientist, AI/ML Engineer, or LLM Engineer roles. Open to hybrid, remote, or on-site.

actively interviewing · response within 24h

Email MeLinkedInGitHubKaggle↓ Resume

Or send a message directly: