AI/ML Engineer focused on production LLM, RAG and recommendation systems. At LikeTik I build hybrid retrieval, LLM re-ranking and multimodal video-analysis pipelines across roughly 6,800 creators, 209,000 videos and 19,500 products — using Qdrant, Whisper, Qwen-VL and self-hosted GPU inference.
Before that I spent three years in industry building churn-prediction pipelines, analytics dashboards and full-stack web applications. I am currently completing an MSc in Artificial Intelligence at THWS Würzburg-Schweinfurt, with a thesis on guardrails and sensitive-data leakage in multi-agent systems.
I care about the parts of AI that make it usable in production: structured outputs that actually validate, retrieval that returns the right thing, and guardrails that catch what should never leave the system.
Quick Facts
LocationWürzburg, Germany
FocusLLM, RAG & Agentic Systems
Experience3+ years in industry
LanguagesGerman (B2) · English (C1) · Hindi & Gujarati (native)
Five-node LangGraph agent pipeline producing source-cited equity research reports across six global exchanges: RAG over SEC 10-K and 10-Q filings via Qdrant hybrid search, real-time news sentiment, and a self-correcting Critic Agent with confidence gating. Hallucination mitigation through grounding, source citations, structured outputs and schema validation; metadata modelled in PostgreSQL (Supabase).
Co-developed a reasoning-tuned Gemma-2B in a team of three for the Google Tunix Hackathon, owning the inference pipeline. Fine-tuned with LoRA (rank 32, alpha 64) on approximately 570k samples from MetaMath, OpenThoughts, Medical-O1, Bespoke-Stratos and GSM8K, with 4-bit NF4 quantization.
Fine-tuned ResNet18 for pneumonia detection on chest X-rays with Grad-CAM and LIME explainability. Compressed a MobileNetV2 variant via structured pruning and INT8 quantization, from 8.7 MB to 4.4 MB — a 49% reduction with minimal accuracy loss.
Trained a YOLOv8 object detector on thermal UAV imagery for automated waterfowl detection, reaching 86.44% mAP@0.5 and 93.21% precision on 83 test images with 1,411 ground-truth annotations. End-to-end CV pipeline from preprocessing to evaluation.
Trained and compared tabular RL agents (Q-Learning, Monte Carlo) that converge toward the mathematically optimal Blackjack strategy. Methods and results documented in a self-authored technical report.
Built a CNN classifier for real-time sign-language gesture recognition across 24 gesture classes, with a preprocessing pipeline and data augmentation for generalisation to unseen hand positions.
Real-time messaging app with bidirectional WebSocket communication and JWT authentication. Room-based architecture deployed on a self-managed Linux server with NGINX.
Security Failure Propagation in Multi-Agent Coding Systems
Om Borda · supervised by Prof. Dr. Ivan Yamshchikov
THWS Würzburg-Schweinfurt · Expected January 2027
Measures how secret and PII leakage propagates across coder, tester and reviewer agents in a LangGraph pipeline, and evaluates a four-layer guardrail ensemble — gitleaks, TruffleHog, Microsoft Presidio and an LLM-as-judge — against GDPR/DSGVO and EU AI Act requirements.
Multi-Agent SystemsLangGraphPII LeakageSecret ScanningGuardrailsLLM-as-JudgeGDPREU AI Act
REPORT
Reinforcement Learning
Optimal Strategy Learning in Blackjack using Reinforcement Learning
Om Borda
2025
A technical report applying reinforcement learning to learn optimal Blackjack strategies. I implement and compare tabular Q-learning and Monte Carlo methods, showing convergence toward near-optimal play after extensive training episodes.
PII DetectionSecret ScanningGuardrailsMicrosoft PresidiogitleaksTruffleHogLLM-as-JudgeGDPR / DSGVOEU AI ActPrompt Injection
AI-Assisted Development
Claude CodeOpenAI CodexCursorGitHub CopilotTask DecompositionCode ReviewTest-Based Validation
Additional Working Knowledge
AWS, Microsoft Azure, GitHub Actions, TensorFlow, MLflow.
Currently Deepening
LLM evaluation and benchmarking with Ragas and Langfuse; tracing and observability; GitLab CI and CI/CD workflows; Model Context Protocol (MCP); LLMOps, model monitoring, drift detection, human-in-the-loop review gates and AI governance.
Built a hybrid recommendation system matching ~19,500 products to ~6,800 analysed creators, combining multilingual embeddings, Qdrant hybrid vector search, metadata filtering and an LLM re-ranking layer returning 5–10 ranked recommendations per creator
Designed the scoring and segmentation layer across eight dimensions — engagement, audience fit, brand fit, reach, discovery, trend fit, cost efficiency and shop conversion — with embedding-based clustering for creator segments and content patterns
Built a multimodal video-analysis pipeline over ~209,000 videos using OpenAI Whisper (Large) for speech-to-text, Qwen-VL for frame-level scene understanding and PyTesseract OCR; extracted signals feed back into the recommendation features
Enforced structured outputs across LLM components with Pydantic response models and JSON Schema, using function calling, tool calling, schema validation and fallback handling on validation failure
Developed production AI APIs with Flask and FastAPI, using MongoDB as the operational database and Qdrant for embeddings, vector search and hybrid retrieval, with batch processing, error handling, logging and data-quality validation
Deployed containerised services with Docker and Kubernetes on Hetzner Cloud, serving a self-hosted Qwen 7B on an NVIDIA A100 40 GB GPU; pipelines designed for ~184,000 creator profiles and more than one million videos
Jun 2023 – Feb 2025 · Part-time alongside studies · Gujarat, India
Built a customer churn prediction pipeline in Python with pandas, scikit-learn and XGBoost, engineering features on activity trends, session duration, failed-payment ratio and support-ticket frequency, with weekly retraining
Queried and processed data from relational client databases using SQL, consolidating structured metadata into the feature pipeline and reporting layer
Developed a sales analytics and forecasting dashboard in Streamlit with KPI cards, regional heatmaps and regression-based forecasts
Full Stack Developer
Greendotslab Software Solutions
May 2022 – Jun 2023 · Part-time alongside studies · Gujarat, India
Built and deployed more than ten client web applications end to end with Node.js, React and REST APIs, with modular architecture and load balancing for production use
March 2025 – January 2027 (expected) · Würzburg, Germany
Master's thesis (in progress): Security Failure Propagation in Multi-Agent Coding Systems, supervised by Prof. Dr. Ivan Yamshchikov — measuring how secret and PII leakage propagates across coder, tester and reviewer agents in a LangGraph pipeline, and evaluating a four-layer guardrail ensemble (gitleaks, TruffleHog, Microsoft Presidio, LLM-as-judge) against GDPR/DSGVO and EU AI Act requirements.
Bachelor of Engineering in Computer Science & Engineering
Gujarat Technological University
July 2020 – May 2024 · Gujarat, India
Grade 9.10/10 (German equivalent: 1.4)
Let's Connect
Available immediately for full-time AI/ML roles in Germany. Interested in collaborating, or have questions about my work? I'd love to hear from you.
