Shardul Pramod Karanjekar
M.Tech Data Science, DIAT Pune (GATE 2024 AIR 1955). Built and shipped ML/NLP systems: fine-tuned FinBERT for financial NER at Infosys, developed a fully local RAG pipeline evaluated on RAGAS, and led generative and computer vision projects using PyTorch and diffusion models. Seeking Data Scientist / ML Engineer roles in applied NLP and GenAI.
Experience
Data Science Intern
Infosys
Remote
Curated a financial NER dataset — 222 documents, 8,652 labeled entities across 11 entity types (ORG, MONEY, DATE, PERCENT, TICKER) — meeting all minimum-instance requirements per class.
Fine-tuned FinBERT on 10-K/10-Q SEC filings using section-aware logic targeting MD&A, Risk Factors, and Financial Statements, improving extraction precision on high-signal regions.
Built a PyTorch + pdfplumber pipeline covering PDF ingestion, section detection, NER inference, and structured JSON output.
Python Expert
Uber AI Solutions
Remote
Designed 15+ LLM evaluation criteria and scored 250+ model outputs for relevance, coherence, and task alignment; feedback directly shaped prompt-engineering and fine-tuning priorities.
Intern
ZCEnergie Solutions
Analyzed 200K+ rows of SCADA data (82 sensors) for anomaly detection; partnered with domain SMEs on FMEA to convert Risk Priority Numbers into turbine maintenance scheduling inputs.
Education
DIAT, Pune
Master of Technology
GPA: 8.42/10 | Coursework: ML, Deep Learning, Computer Vision, Data Structures, Descriptive Statistics
PVG College of Engineering, Pune
B.E.
GPA: 8.07/10 | Coursework: IoT, Electric Vehicles, Power Electronics
Skills
Projects
CACGen: Adaptive Camouflage Image Synthesis via ControlNet Diffusion — M.Tech Thesis, DIAT
Cut KID to 0.0081 on the camouflaged split (62% lower than LAKE-RED) using a mode-aware prompt builder that switches between background pixel analysis and CLIP-based habitat inference, removing foreground leakage.
Outperformed LAKE-RED across all 3 evaluation splits (CMMD 0.0449) with a TELEA-based adaptive control image builder; improved CamOT in 63.1% of 6,495 COD10K test cases.
Tax RAG: Local Retrieval-Augmented Generation for Indian Tax Compliance [Link]
Built a fully offline RAG pipeline (Gemma 4B GGUF + FAISS) indexing Income Tax Act sections and government PDFs with zero API dependency, suited for privacy-sensitive compliance use.
Scored 0.99 faithfulness, 0.90 answer relevancy, 0.84 context recall, 0.65 context precision on 25 held-out queries via RAGAS with bge-base-en-v1.5 embeddings; flagged precision as the key gap.
Olist E-commerce Analytics & Prediction System [Link]
Surfaced +7.1% median MoM revenue growth, cohort retention, seller ranking, and delivery-delay trends via a SQL analytics layer (PostgreSQL, window functions, CTEs) over 96K+ orders ($15.4M revenue).
Reached PR-AUC 0.89 with an XGBoost churn model on RFM and delivery/payment features (SMOTE, SHAP); identified delivery delay and Boleto payments as top churn drivers.
Shipped a FastAPI service with a Bootstrap + Plotly dashboard serving live KPIs and cohort charts.
Industrial Computer Vision Portfolio [Anomaly detection] [Steel defect detection]
Achieved 0.9824 mean Image AUROC / 0.9801 Pixel AUROC with PatchCore (WideResNet50, FAISS memory bank) across all 15 MVTec AD categories, matching published SOTA.
Reached mAP@50 0.7811 with YOLOv8s on NEU steel-defect data via CutPaste-Poisson augmentation, beating ControlNet-generated augmentation (0.7493); logged LoRA fine-tuning as a negative result.
Sped up YOLOv8 inference 1.71× (102 → 174 FPS) by deploying via TensorRT FP16 on RTX 4050 (6GB VRAM).
Drug Consumption Prediction using Machine Learning
Reached 87% accuracy (>90% on Cannabis, Cocaine, Ecstasy) with an XGBoost/RF/SVM/MLP ensemble on 1,885 UCI Drug Consumption samples across 19 substances.
IoT Energy Monitoring System — B.E. Final Year Project
Built an ESP32-based energy monitoring system (ZMPT101B/ACS712 sensors) streaming real-time voltage, current, and power to Blynk Cloud with remote load control.