cv

Software Developer

Medical Informatics Engineering

Sole developer of WorkWell Measure Studio, a clinical quality measure platform for WebChart. Retired its Java/Spring Boot backend in favor of TypeScript across 500+ pull requests, with 3,268 automated tests.

• Runs CMS's own FHIR measure artifacts behind a permanent CI gate: all 455 official MADiE test cases across nine measures pass, and results agree with HAPI's independent engine on 362 of 387 cases.
• Scaled a 20,000-patient sandbox to 120,000 nightly evaluations. Batching made official execution 10 to 16x faster and cut a CSV export that was timing out at 60 s to 7.7 s.
• Integrated live HL7 FHIR R4 data over SMART Backend Services and shipped a 13-tool read-only MCP server, CDS Hooks cards, and MeasureReport and QRDA exports that validate with zero errors.

Product Associate, Operations and Analytics

Riccle (early-stage e-commerce startup)

Ran operations reporting and analytics for a small e-commerce team.

• Merged order, inventory, payment, and customer data from four separate trackers into one source and built recurring KPI reports with SQL and Python, saving about six hours a week.

Graduate Researcher, AI Safety and LLM Evaluation

Purdue University Fort Wayne

Studied how LoRA, QLoRA, 4-bit quantization, and full fine-tuning change jailbreak robustness in open-weight LLMs, advised by Dr. Jonathan Rusert.

• Fine-tuned and evaluated 25 LLM configurations across five model families (4B to 14B parameters) with SLURM on A100 and A30 GPUs.
• Used a pre-specified ±5 pp equivalence test (TOST) to show that LoRA and full fine-tuning degrade black-box jailbreak safety equivalently under a controlled recipe, across nearly 600 attack-success measurements.
• Ran a 750-sample blind human annotation study. GPT-4o-mini mislabeled 45.7% of safe outputs as jailbreaks, against 8.6% for a calibrated classifier, which inflated measured degradation by about 3x.
• For the JUDGe study, showed that GPT-5.6 judges still over-flagged 21.5–29.9% of safe completions and gave identical outputs different labels on 3–15% of repeat evaluations.

Research Collaborator, Query-Efficient Adversarial NLP (Dr. Jonathan Rusert's group)

Purdue University Fort Wayne

Studied where black-box text attacks such as TextFooler and TextBugger spend their query budget, separating word selection from replacement and comparing datasets, victim models, and attack traces to find where fewer queries would do the same job.

Software Development Intern, NLP and Retrieval-Augmented Generation

Medical Informatics Engineering

Built and containerized the first prototype of RASS, a RAG platform for semantic search over 3,000+ Redmine issues, with hybrid search, reranking, and MCP and SSE interfaces.

• Benchmarked relevance, grounding, and latency across retriever and model settings, and load-tested the service at 25 concurrent users and AI agents with responses under 500 ms.
• Added OpenTelemetry tracing so retrieval and service failures were easy to diagnose.

M.S.

Purdue University Fort Wayne

Computer Science

• Master's thesis: Do Efficient Adaptations Reduce Safety? Jailbreak Robustness of PEFT vs. Full Fine-Tuning on Consumer-Accessible LLMs
• Thesis advisor: Dr. Jonathan Rusert

B.S.

Lahore University of Management Sciences

Computer Science

Your Judge Is a Confound: Evaluator and Attack Choice Distort Jailbreak-Safety Measurement for Fine-Tuned LLMs

First Workshop on Reliable Evaluation for Language Models (JUDGe), NeurIPS 2026

Taleef Tamsal and Jonathan Rusert. Accepted for oral presentation (non-archival); also accepted at AdvML-Frontiers × CoTMA, COLM 2026.

Adaptation, Not Algorithm: LoRA and Full Fine-Tuning Show Comparable Black-Box Jailbreak Degradation in Five Open-Weight LLMs

Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026 (accepted)

Taleef Tamsal and Jonathan Rusert. Peer-reviewed archival paper.

PFW at SemEval-2026 Task 6: Multi-Seed DeBERTa Ensembles for Political Response Clarity and Evasion Classification

Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval 2026), pp. 1518–1525, ACL

Taleef Tamsal and Jonathan Rusert. Ranked 18/41 on clarity and 12/33 on evasion.

PFW Task 8 at SemEval-2026 Task 8: Lightweight Tri-Fusion Retrieval with Prompt-Engineered Faithful Generation for Multi-Turn RAG

Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval 2026), pp. 1526–1532, ACL

Taleef Tamsal and Jonathan Rusert. Ranked 6/26 on Task B.

Do Efficient Adaptations Reduce Safety? Jailbreak Robustness of PEFT vs. Full Fine-Tuning on Consumer-Accessible LLMs

M.S. thesis, Purdue University Fort Wayne

Master's thesis on adaptation methods, jailbreak robustness, evaluator disagreement, and safety-efficiency trade-offs in consumer-accessible LLMs.

Presenter

ACL 2026 (virtual)

Presented the SemEval 2026 Task 6 and Task 8 papers.

Reviewer

SemEval 2026 Tasks 6 & 8; JUDGe 2026 (NeurIPS Workshop)

Mentor

GradApp Lab; LUMS Alumni Mentorship Program

Teaching Assistant

Lahore University of Management Sciences

Algorithms (CS 310), Introduction to Blockchain (CS 3812), Topics in Blockchain Technology (CS 589), and Computational Problem Solving (CS 100).

JobOps Copilot, multi-tenant AI agent platform

Next.js, TypeScript, and Python platform that runs stateful LangGraph agents with pgvector RAG and MCP tools, deployed on Azure with a live demo.

• Tenant isolation and PII redaction
• CI/CD quality gates
• Job-ranking correlation raised from 0.716 to 0.821

CorvEd

Full-stack tutoring operations platform covering intake, payments, matching, and scheduling.

• Next.js App Router
• Supabase Postgres
• Row Level Security
• PKCE OAuth

WorkWell Measure Studio

Production healthcare compliance platform with deterministic CQL evaluation, interoperability, and guarded AI assistance.

• TypeScript
• FHIR R4
• SMART Backend Services
• CQL
• MCP
• PostgreSQL

TurboQuant

PyTorch implementation and evaluation of KV-cache compression for long-context LLM inference.

• 5.2x validated cache compression
• Qwen retrieval-safe evaluation through 32K context
• PyTorch
• Hugging Face DynamicCache

RASS / CoRAG

Multi-service retrieval platform with asynchronous ingestion, hybrid search, grounded streaming answers, citations, and observability.

• Node.js
• OpenSearch
• Redis
• PostgreSQL
• Docker
• MCP and SSE

UNLP 2026 Ukrainian Document QA

Retrieval-augmented question answering pipeline combining BGE-M3 retrieval, Qwen3 reranking, and Gemma-based multiple-choice answering.

• Private score 0.8722
• Ranked 7/15
• BGE-M3
• FAISS
• Qwen3