Selected WorkSystems That Shipped
Eight real projects across AI/ML, MLOps, and full-stack — each with the problem it solved.
PRJ-01status: shipped
AI/MLTecholution
Production translation that actually respects domain terminology.
English-to-Dutch Translation System
problem
Generic translation models mangled domain-specific glossary terms and drifted over time.
solution
Built a Fusion-RAG pipeline with a custom glossary and a continuous MLOps retraining loop to keep quality climbing.
▸ 12pt BLEU increase · 30% accuracy gain
LLMs · Fusion RAG · Fine-Tuning · MLflow · Airflow
LLMs → Fusion RAG → Fine-Tuning → MLflow → Airflow
PRJ-02status: shipped
AI/MLTecholution
Transcription that finally gets Indian-origin words right.
Whisper Fine-Tuning System
problem
Off-the-shelf Whisper mis-transcribed Indian names and domain vocabulary, breaking downstream NLP.
solution
Fine-tuned Whisper with a custom vocabulary and curated audio set, then served it behind a fast inference path.
▸ 50% accuracy improvement
Whisper · Fine-Tuning · Python · PyTorch
Whisper → Fine-Tuning → Python → PyTorch
PRJ-03status: shipped
AI/MLTecholution
An agent platform with no framework lock-in — full control of the loop.
Centralized RLEF Assistant
problem
Off-the-shelf agent frameworks hid the control flow and made debugging production behavior painful.
solution
Built custom agents and tools from scratch — orchestration, tool-calling, and reinforcement-from-feedback loop, all owned in-house.
▸ Zero external agent frameworks
AI Agents · LLMs · Python · FastAPI
AI Agents → LLMs → Python → FastAPI
PRJ-04status: shipped
AI/MLCulinda
Turned a sprawling security knowledge base into instant answers.
Cybersecurity RAG Chatbot
problem
Security teams burned hours hunting through scattered documentation to resolve recurring issues.
solution
Built an NLP + LLM RAG chatbot grounded in the internal knowledge base with citation-backed responses.
▸ 30% issue-resolution improvement
RAG · NLP · LLMs · Vector DB
RAG → NLP → LLMs → Vector DB
PRJ-05status: shipped
Full-StackCulinda
Let non-technical users query data in plain English.
NLP-to-SQL Electron App
problem
Analysts needed SQL fluency to explore data and build visualizations — a hard bottleneck.
solution
Shipped a desktop Electron app on a fine-tuned NLP-to-SQL model with natural-language-editable data visualizations.
NLP-to-SQL · Fine-Tuning · Electron · React
NLP-to-SQL → Fine-Tuning → Electron → React
PRJ-06status: shipped
MLOps
Model retraining that runs itself, end to end.
Automated CI-CD ML Pipeline
problem
Manual retraining and deployment slowed iteration and invited human error.
solution
Built an Airflow + MLflow pipeline on AWS EC2 for a Medical MNIST classifier — automated training, tracking, and deployment.
▸ 30% faster model retraining
Airflow · MLflow · AWS EC2 · Docker · CI/CD
Airflow → MLflow → AWS EC2 → Docker → CI/CD
PRJ-07status: shipped
AI/MLVzya
A research bench for ML-driven trading ideas.
AI Trading Strategy Platform
problem
Trading hypotheses needed rigorous, repeatable backtesting before any capital was at risk.
solution
Researched, implemented, and backtested 15+ ML trading strategies with a unified evaluation harness.
▸ 15+ strategies backtested
Python · ML · Backtesting · Pandas
Python → ML → Backtesting → Pandas
PRJ-08status: shipped
AI/ML
Wearable assistance, controlled entirely by voice.
AI Glasses for Visually Impaired
problem
Visually impaired users needed hands-free, real-time help understanding their surroundings.
solution
Built smart glasses with speech recognition driving 4+ voice-controlled functionalities for daily assistance.
▸ 4+ voice-controlled functions
Speech Recognition · Computer Vision · Python · Embedded
Speech Recognition → Computer Vision → Python → Embedded
