surface view
GPT-2 BPE TOKENIZER  ·  VOCAB_SIZE: 50,257  ·  src: bio.txt
token stream →
DRAFT  v1.0
arxiv: cs.CL/2025.aryasuneesh-personal Under Continuous Revision

Arya Suneesh

NLP Engineer AI Researcher Quantum Enthusiast Human-Centered AI
Abstract

Building systems that read between the lines. Interested in how language encodes meaning, how attention mechanisms mirror human cognition, and what happens at the boundary between classical and quantum computation. Equally drawn to human-centered AI — building tools that adapt to people rather than demanding people adapt to tools. One of 16 selected globally for the MCQST Summer Bachelor Program; DAAD-WISE Scholar; 3rd place at SemEval 2025 & CLEF CheckThat! 2025.

ENTITIES → PERSON ORG GPE TECH hover any entity to reveal label
§ 1

Biography

revised
May 2026 Arya Suneesh is an NLP researcher and engineer whose work spans language, finance, and quantum computing. At TIFIN RM she built semantic caching layers and LLM agents for wealth management intelligence, reducing query latency by 40–70% across 57 financial query types. She completed her B.Tech. (Hons.) in Computer Science at IIIT Kottayam with a GPA of 9.64/10, and was selected as one of 16 participants worldwide for the MCQST Summer Bachelor Program at the Walther-Meißner-Institut in Munich. GPA verified
May 2025
A thread running through her work is a human-centered approach to AI — building systems that adapt to people rather than the reverse. This came into sharpest focus during her DAAD-WISE residency at Universität der Künste Berlin and the Einstein Center Digital Future, where she worked in human-computer interaction (HCI), developing ASR models for the Reading Primer — a personalised learning device that meets readers where they are. She also serves as Community Lead, ML Industry at Cohere Labs Open Science Community. HCI thread
explicit now
◈ DEPENDENCY PARSE · compromise.js · en (interactive, client-side)
loading NLP library…
nsubj  ·  dobj  ·  relcl  ·  prep  ·  pobj  ·  amod   |   purple arc = cross-clause   |   max 12 tokens
◈ CONCEPT EMBEDDING · UMAP(n_neighbors=15, min_dist=0.1) · research domains
RAGNERASRLLMTransformersfine-tuningNLPSem. CacheFinTechFinancial NLPFinanceVQEQMLQCNNshot-noisebarren plateauqubitsQuantumReading PrimerHCIudk.aiLearning TechHCINLP↔FinanceNLP↔HCINLP↔Quantum
each point = a research concept · cluster colour matches role-tag legend · dashed edges = cross-domain bridges in Arya's work
§ 2

Experience

[COMPLETED]
NLP Engineer — TIFIN RM · Bengaluru
May 2025 – Jan 2026
  • Engineered a semantic caching layer using MPNet embeddings; reduced response latency by 40–70% across 57 financial query types in the client diagnostics module.
  • Optimized LLM agents for financial advisory workflows via prompt engineering, data filtering, and token budgeting; extended SIP/STP/SWP coverage with fault-tolerant fallback logic.
  • Built anonymization pipelines and evaluation frameworks; resolved critical production bugs in context routing and response generation.
add eval
benchmarks
[COMPLETED]
Student Researcher — Walther-Meißner-Institut (WMI) · Munich
Jul 2025 – Aug 2025  ·  MCQST Summer Bachelor Program  ·  1 of 16 worldwide
  • Developed a shot-based simulation framework for realistic quantum measurements with post-selection to preserve gauge symmetry; designed for a 17-qubit superconducting chip.
  • Characterized VQE optimization landscapes across shot-noise levels; quantified the impact of measurement statistics on convergence behavior.
  • Contributed to experiment design and algorithm architecture within an international team of six researchers.
[COMPLETED]
AI Research Intern — TIFIN RM · Bengaluru
Jan 2025 – Apr 2025
  • Built semantic caching system for the client diagnostic tool (~60% latency reduction across 57 query types).
  • Improved LLM response quality via data-filter rules and structured prompt variants; diagnosed jq command inconsistencies causing latency spikes in advisor portfolios with 10,000+ clients.
◈ LATENCY BENCHMARK · semantic caching (MPNet embeddings) vs. baseline · client diagnostic module
baselinecachedClient Diagnostics−60%SIP Advisory−65%SWP Queries−70%Portfolio Overview−55%Advisory Q&A−42%baseline →
40–70% response-latency reduction across 57 financial query types · baseline = uncached LLM call · values representative
[COMPLETED]
AI Research Intern — MyFi by TIFIN · Bengaluru
Oct 2024 – Dec 2024
  • Benchmarked conversational models for a mobile financial assistant; evaluated quality-latency tradeoffs for production deployment.
  • Integrated Indian language and colloquial input support into the MyFi stack.
[COMPLETED]
Research Intern (DAAD-WISE Scholar) — UdK Berlin / ECDF · Berlin
May 2024 – Jul 2024  ·  DAAD-WISE Scholarship  ·  1 of 250 selected from India  ·  Human-Centered Data Science · HCI
  • Worked at the intersection of human-computer interaction and human-centered data science under Prof. Daniel D. Hromada; built NLP components that adapt to individual readers rather than imposing fixed curricula.
  • Developed speech recognition components and evaluated ASR models for on-device learning modules in the Reading Primer device; iteratively advanced the Fibel 4 prototype across multiple hardware-software cycles.
  • Adapted personalised NLP models from udk.ai and validated their integration into the Reading Primer pipeline.
◈ SELF-ATTENTION · BERT-base layer 8 head 5 · "attention maps meaning across token boundaries"
query (row) → key (col) · cell opacity = attention weight · diagonal = self-attention · darker = higher weight
§ 3

Selected Projects

[ACTIVE]
Reading Primer
UdK Berlin / ECDF · DAAD-WISE · 2024
An NLP-powered reading companion that adapts difficulty and surfaces connections across texts — designed for on-device, low-latency learning. Core insight: reading comprehension is a graph traversal problem; every unfamiliar word is an edge to a new subgraph. Built ASR and personalised NLP components integrated into the udk.ai pipeline.
primary
project
[PUBLISHED]
QCNN-MFND — Quantum CNN for Multimodal Fake News Detection
Honours Thesis · IIIT Kottayam · 2025  ·  QNLP @ AACL-IJCNLP 2025, Mumbai
A novel quantum convolutional neural network framework for detecting fake news across text and image modalities in social media. Combines classical multimodal encoders with parameterised quantum circuits to exploit entanglement for cross-modal feature fusion. Supervised by Dr. Balasubramanian P.
[SEE §5]
Quantum VQE  Variational Quantum Eigensolver
Walther-Meißner-Institut · MCQST · 2025
Shot-based simulation and optimisation of VQE circuits on a 17-qubit superconducting processor. Investigation of barren plateau phenomena under realistic shot-noise conditions. Won't explain further here → full write-up in §5 (access required).
§ 4

Publications

Suneesh, A., Balasubramanian P. "QCNN-MFND: A Novel Quantum CNN Framework for Multimodal Fake News Detection in Social Media." QNLP Workshop, AACL-IJCNLP 2025, Mumbai. Hazarika, B., Suneesh, A., Devadiga, P., Rajpoot, P.K., Suresh, A.B., Hussain, A.I. "Multilingual Conversational AI for Financial Assistance: Bridging Language Barriers in Indian FinTech." FinAI Workshop, CIKM 2025, Seoul. Sharma, M., Suneesh, A., Jain, M., Rajpoot, P.K., Devadiga, P. et al. "TIFIN at CheckThat! 2025: Reasoning-Guided Claim Normalization for Noisy Multilingual Social Media Posts." CheckThat! Workshop, CLEF 2025, Barcelona. 3rd place · Task 2 Gurumurthy, K., Shrivastava, A., Rajpoot, P.K., Devadiga, P., Hazarika, B., Jain, M., Sharma, M., Suneesh, A. et al. "TIFIN at CheckThat! 2025: Cross-Lingual Subjectivity Classification in News through Monolingual, Multilingual, and Zero-Shot Learning." CheckThat! Workshop, CLEF 2025, Barcelona. Hazarika, B., Devadiga, P., Rajpoot, P.K., Baliga, A.U., Gurumurthy, K., Jain, M., Sharma, M., Shrivastava, A., Suneesh, A., Suresh, A.B. "TIFIN at CheckThat! 2025: X-VERIFY – Multi-lingual NLI-based Fact Checking with Condensed Evidence." CheckThat! Workshop, CLEF 2025, Barcelona. Devadiga, P., Suneesh, A., Rajpoot, P.K., Hazarika, B., Baliga, A.U. "TIFIN India at SemEval-2025: Harnessing Translation to Overcome Multilingual IR Challenges in Fact-Checked Claim Retrieval." SemEval 2025, ACL 2025, Vienna. 3rd place · Task 7
§ 5

██████████ ██████████

⚠  CLASSIFIED CLEARANCE LEVEL: QUANTUM-1

████████████████████████████████████████████████████████████████████ ████████████████████████████ ███████████ ██████████████████████████ Research conducted at Walther-Meißner-Institut, Munich under the MCQST Summer Bachelor Program using shot-based VQE simulation on a 17-qubit superconducting processor.

Research conducted at Walther-Meißner-Institut (WMI) in Munich under the MCQST Summer Bachelor Program (1 of 16 selected worldwide). Developed a shot-based simulation framework for realistic quantum measurements with post-selection to preserve gauge symmetry, designed for deployment on a 17-qubit superconducting chip.
Key finding: VQE optimization landscapes exhibit the barren plateau phenomenon even under finite shot-noise — vanishing gradients that emerge in high-dimensional Hilbert spaces are compounded by measurement statistics in ways the noiseless theory obscures. The boundary between classical and quantum ML is more porous than either field admits. key finding —
write-up pending
Characterised convergence behaviour across varying shot-noise levels in collaboration with an international team of six researchers. Contributed to experiment design and algorithm architecture for the WMI 17-qubit processor.
◈ VQE PARAMETER LANDSCAPE · θ₁ × θ₂ sweep [0, 2π] · ⟨E⟩ under finite shot-noise
dark = low energy (minimum) · flat plateau = vanishing gradients → barren plateau · gradient ≈ 0 for >90% of parameter space
🔑 access requires quantum key:
aryasuneesh.github.io  ·  living document  ·  continuously revised computing token count…