azeebneuron
Work
  • Apr2026 - Present
    Blinkit
    Data Scientist
    • Recently joined; currently ramping up on the team’s GenAI workstreams.
  • Dec2025 - Feb2026
    LastDraft AI
    Founding Engineer
    • Built a summary-first extraction pipeline (tender metadata in under 5 minutes), RAG-based semantic search over 300+ page documents, and a hybrid bid eligibility engine using policy thresholds, fuzzy region matching, and custom LLM rules for bid/no-bid decisions.
    • Implemented a two-phase BOQ extraction pipeline (detect → per-chunk extract) with multi-BOQ evaluation framework; engineered a DOCX form-fill engine for automated bid document generation.
    • Built a ReAct-style BidDraftingAgent with 4 specialized tools (pgvector tender search, vault RAG, past bids retrieval, company metadata) that autonomously drafts legally-compliant bid documents, cutting preparation time from days to under an hour.
    • Owned end-to-end AWS deployment (EC2, RDS, S3) via Docker Compose; implemented Google OAuth + bcrypt/JWT auth, brute-force rate limiting, XSS sanitization, and idempotent SQL migrations for zero-downtime schema rollouts.
  • Mar2025 - Dec2025
    Docxster
    Machine Learning Intern → Engineer
    • Engineered a synthetic data generation pipeline using a GAN to produce a custom dataset of complex document images, addressing the lack of real-world training data.
    • Architected and developed DocStruct-YOLO, a YOLOv10-based model for Document Layout Analysis (DLA), precisely identifying tables, text, and figures as the foundational step for a proprietary extraction module.
    • Built a cost-efficient document processing module integrating Visual Language Models (VLMs) with OCR, achieving Textract-level accuracy while reducing inference costs by 40% across varied document types.
  • Jul2024 - Aug2024
    Sudha Gopalakrishnan Brain Centre, IIT Madras
    Machine Learning Intern
    • Developed a Python pipeline to convert SVG brain annotations into validated GeoJSON format, utilizing geospatial libraries to enable accurate mapping and interoperability with advanced analysis tools.
    • Fine-tuned ResNet-50 for brain region similarity analysis, generating quantitative similarity scores that reduced manual review time by 10 hours/week and accelerated the research pipeline.