87%
MIRAGE benchmark accuracy
Built an agentic oncology intelligence system with an evidence-weighted knowledge graph and auditable retrieval, 7.6 points above baseline.
- ~18KGraph entities
- ~95KRelationships
Applied Scientist · NLP, Retrieval & LLM EvaluationNew York, NY
Reliable ML systems,built for real-world data.
Applied scientist and ML engineer building retrieval, classification, and LLM evaluation systems. I pair production ownership with rigorous experimentation, model validation, and research on robustness and efficiency.
Selected outcomes
Measured results across production ML, retrieval, knowledge systems, and model safety.
~60%
Less manual correction time
Built a schema-validated document-intelligence pipeline for scanned PDFs, forms, and tables across 800 documents.
- 95%Exact field accuracy
- 5KLabeled fields
16×+
More intervention required to break refusal
Developed a training-time safety defense for an open-weight LLM while preserving baseline behavior.
- 1 directionBefore defense
- 16+ directionsAfter defense
~1M
LLM responses processed
Built the evaluation backend for a production measurement platform spanning six LLMs and more than 10,000 prompts each week.
- ~93%Intent accuracy
- <2 hrsRecurring analysis
01
What I work on
Production ownership, careful evaluation, and research depth across the ML lifecycle.
01Systems
Retrieval systems built for production
I build hybrid retrieval, ranking, and multi-agent systems with traceable evidence, calibrated routing, and latency and cost measured against real constraints.
02Evaluation
Measurements that hold up on real data
I design benchmarks, confidence thresholds, and out-of-domain checks that expose failure modes and make model quality useful for operational decisions.
03Research
Safer, more efficient language models
I study how model behavior is represented, attacked, and improved. Current work spans refusal robustness, activation-level interpretability, and tokenizer optimization.
02
Professional experience
Production ML work that improved quality, latency, and the decisions teams could make.
Jun 2025 - Aug 2025
87%MIRAGE accuracy
~18KGraph entities
~95KRelationships
Statistics & AI/ML InternPfizer · Boston, MA
- Built an AWS-based agentic oncology intelligence system that let researchers query emerging drug-disease-target relationships extracted from scientific conference abstracts, achieving 87% accuracy on the MIRAGE benchmark, 7.6 points above baseline.
- Designed ontology-driven normalization using HGNC and DOID to standardize biomedical entities and identifiers across literature sources.
- Built a Neo4j knowledge graph with approximately 18,000 entities and 95,000 evidence-weighted relationships, each linked to canonical identifiers, source text, and provenance.
- Integrated eight retrieval tools into a multi-agent LangGraph workflow with task-aware routing between Gemini Flash and DeepSeek-R1, so answers returned auditable evidence paths through planning, graph traversal, validation, and synthesis.
Jul 2023 - Jul 2024
84%Top-3 exact match
92%Answer F1
900 msp95 retrieval
Data ScientistIncedo Inc. · Gurugram, India
- Productionized a hybrid retrieval stack with BM25, dense retrieval, and cross-encoder reranking across 1,200+ technical manuals for a telecom router company. Top-3 exact match rose from 41% to 84%, answer F1 from 68% to 92%, and p95 latency fell from 2.3 seconds to 900 milliseconds.
- Developed backend components for a customer-facing document-intelligence pipeline that converted scanned PDFs, forms, and tables into schema-validated structured records, achieving 95% exact field accuracy across 5,000 labeled fields from 800 documents and reducing manual data-entry and correction time by approximately 60%.
- Fine-tuned open-weight LLMs for medical and telecom named-entity recognition, improving micro-F1 from 0.75 zero-shot to 0.90 on a 2,500-example in-domain test set.
- Measured out-of-domain decay to 0.61 F1 and used the analysis to define when fine-tuning outperformed prompt-based extraction.
- Deployed document QA and multimodal RAG using FAISS on Azure ML with evaluation guardrails for relevance, citation support, factual consistency, completeness, and jailbreak behavior. The suite reached 87% precision and 82% recall, routing 12% of outputs to human review.
Jan 2023 - Jun 2023
RISC-VVector extension
LLVMCustom lowering
Software Engineering InternKinara AI (acquired by NXP Semiconductors) · Hyderabad, India
- Prototyped a RISC-V vector extension and LLVM backend with custom lowering for vectorized memory and scatter-gather operations, accelerating ML-kernel execution in cycle-accurate simulation.
Feb 2026 - May 2026
~1MLLM responses
~93%Intent accuracy
<2 hrsRecurring analysis
Data Analyst InternLiving Brands AI · Brooklyn, NY
- Built the evaluation backend for a production platform that lets brand-strategy teams benchmark prominence across six LLMs and more than 10,000 weekly prompts, tracking first-mention rank, share of voice, sentiment, and citation coverage.
- Developed the statistical methodology for the platform's cross-brand comparisons, so it reports one brand as surfacing above another only when the difference across sampled prompts is statistically significant rather than sampling noise.
- Processed approximately 1 million responses through an embedding-retrieval and two-stage classifier, reaching approximately 93% intent accuracy and routing ambiguous mentions to human review.
- Delivered React and TypeScript dashboards for weekly client reporting, cutting recurring cross-model analysis from approximately two days to under two hours. The AWS-hosted system remained in production after the internship.
03
Selected work
Three focused examples of applied systems and research, each framed by the problem, method, and result.
16×+ harderTo remove refusal behavior
Baseline model behavior preservedRobustness & Evaluation of Refusal in Open-Weight LLMs
Developed a training-time defense and evaluation harness for an open-weight LLM, then stress-tested whether refusal behavior could be removed through low-rank linear ablations.
Increased the rank needed to break refusal from 1 to at least 16 while preserving baseline model behavior.
0.78% fewer tokensThan BPE at matched vocabulary size
Near-optimal solutions with GurobiJoint Optimization for Greedy Longest-match Tokenization (JOLT)
Developed a Gurobi mixed-integer optimization approach for tokenizer vocabulary selection that matches the greedy decoding rule used by WordPiece.
Reduced token count by up to 0.78% versus BPE at the same vocabulary size and closed 89.6-99.4% of the remaining compression gap. Accepted at COLM 2026 Tokshop.
Top 8 / 30Dell × NVIDIA Hackathon
Fully local vision-language inferencePhysioCoach: Real-Time AI Physical-Therapy Coach
Built a fully local physical-therapy coach that combines Qwen2.5-VL form feedback, MediaPipe repetition tracking, range-of-motion analysis, and spoken guidance in the browser.
Top-8 of 30 teams at the Dell × NVIDIA Hackathon 2026 (NYU CDS).
Additional projects4 more
Input-Only Suppression of Evaluation-Awareness Latents
Consistently suppressed targeted activations, including fully turning off one causally validated feature, but found that readable activations do not always control behavior.
Imbalanced Binary Classification on Graph-Structured Data (RNA GNNs)
Improved ROC-AUC by 6.2 percentage points over CNN baselines and shipped an interactive 3D demo.
World-Model Training & Forecasting for Robot Locomotion
Evaluated motion forecasts across 1.6, 3.2, and 5.0 second horizons using deployment-focused trajectory metrics instead of single-step error alone.
Manuscript in preparationAxentra & LaRa-Home · NYU Leslie eLab Startup Bootcamp
Ran customer discovery and MVP scoping to validate pain points across clinical research operations and architecture / building-code review.
04
Research & credentials
Selected publications, education, and the technical foundation behind the work.
Selected publications
Joint Optimization for Greedy Longest-match Tokenization
Developed a mixed-integer optimization approach that matches the greedy decoding rule used at deployment, reducing token count by up to 0.78% versus BPE at the same vocabulary size.
Accepted · COLM 2026 Tokshop · Co-first authorMinimizing Targeted Activations: Input-Only Suppression of Evaluation-Awareness Latents in Large Language Models
Showed that fluent prompts can suppress five types of internal model features without changing model weights, while control experiments revealed that readable activations do not always control behavior.
Preprint · arXiv · 2026Validity of Machine Learning-Based COVID-19 Prediction
Validated seven classifiers on 195,000 clinical records, measured approximately 20% AUROC degradation under cross-continental distribution shift, and released an open-source evaluation toolkit.
Earlier publications3 more
Auto Encoders for Communication-Efficient Distributed Learning
Workshop paper · AAAI Deployable AI · 2023
AI in Coronary Physiology: Where Do We Stand?
Review article
Exploring Protein Design Landscapes with Semi-Supervised Adaptive Sampling
22nd Int'l Conference on Bioinformatics, Brisbane · Nov 2023
Building AI that needs to work beyond the demo?
I am open to applied scientist, machine learning engineer, data scientist, and research engineer roles.
dm6262@nyu.edu