About Me

Ayush Dongre Machine Learning Engineer

I am an Integrated M.Sc. (5-Year) in Mathematics & Computing graduate from the Indian Institute of Technology, Kharagpur (2019 — 2024).

Experience

Architected a serverless RAG pipeline, minimizing idle infrastructure costs via on-demand scaling of Dockerized microservices.
Reduced indexing costs by 90% relocating segmentation to local SLMs (Qwen 3 : 0.6B) & using SOTA APIs for entity extraction.
Implemented “Ancestry Injection”, embedding document hierarchies, solving context fragmentation & lost in the middle problem.
Replaced imprecise chunking with perplexity-based segmentation using dynamic log-loss thresholds to split at narrative breaks.
Integrated KuzuDB to enable O(1) adjacency traversals & complex Cypher queries, replacing unbundled Parquet-based storage.

Evaluated models using Python & C++ red-teaming, ensuring strict instruction-following and factual correctness confirmations.
Built SFT datasets with GitHub repos to fine-tune Gemini 1.5 Pro for code, ensuring robustness via Docker-based unit testing.
Developed Python pipelines to generate 200+ complex data visualisations that trained Gemini’s multimodal capabilities efficiently.
Audited models via RLHF on 2K+ math & 200+ coding problems, finding logical & stylistic flaws to improve response reliability.

Engineered NLP classification pipelines to automate high-volume ticket routing, greatly reducing manual triage complexity.
Fine-tuned MiniLM models via UMAP & HDBSCAN clustering to automate labeling, achieving sustained 85% balanced accuracy.
Migrated legacy ETL to production-grade Python pipeline, optimizing logic to reduce manual approval intervention.