I work across the complete AI product lifecycle:
data → models → evaluation → APIs → production.
|
Founding Engineer @ Uniiq.ai
May 2026 – Present · Production AI, Backend Systems & Product Engineering
|
|
ML Researcher @ Yale University
October 2025 – March 2026 · Secure Deep Learning & Compiler-Centric MPC
|
|
Quantitative Research Consultant @ WorldQuant BRAIN
November 2022 – July 2025 · Alpha Research & Workflow Automation
|
| AI & Machine Learning |
|
| Speech & NLP |
|
| Backend & Data |
|
| Product Engineering |
|
| MLOps & Infrastructure |
|
| Languages |
|
Natural-language analytics over a transactional knowledge graph.
- Transformed 21,393 transactional records across 19 relational tables into a knowledge graph containing 765 nodes and 877 edges.
- Implemented a two-stage LLM orchestration pipeline with query classification, validation, graph traversal, and answer generation.
- Delivered the system through a FastAPI backend and React interface with real-time streaming using Server-Sent Events.
End-to-end Hindi speech-recognition training, normalization, and evaluation pipeline.
- Fine-tuned Whisper-small using SpecAugment, label smoothing, and cosine learning-rate scheduling.
- Built a Devanagari normalization engine with phonetic reverse transliteration to handle spelling variation and loanwords.
- Combined five ASR outputs through word-level consensus evaluation and achieved a documented greater than 48-percentage-point absolute WER reduction against the stated baseline.
Privacy-preserving collaborative AI research at Yale University.
- Developed deep-learning components from first principles without an external autograd framework.
- Worked on compiler-centric MPC architecture and contributed to the open-source
sequreecosystem.- Designed a custom CNN that achieved 88.08% test accuracy on ChestMNIST in the documented evaluation.
Deep-learning experiments for particle-physics problems in the ML4SCI ecosystem.
- Developed ResNet-style and custom CNN architectures for photon identification and quark/gluon classification.
- Explored graph-neural-network architectures for momentum regression on high-dimensional detector data.
- Compared GCN and GAT approaches across accuracy, complexity, and latency considerations.
Retrieval-augmented generation for biomedical question answering.
Combined biomedical embeddings, Qdrant vector retrieval, and domain-focused generation to build a specialized scientific question-answering pipeline.
