Deep Learning · LLM EngineeringGPT from Scratch - PyTorch
Using pretrained models alone does not provide first-principles understanding of tokenization, attention, training dynamics, and autoregressive generation.
PyTorchCustom BPE TokenizerTransformerMulti-Head Attention
RAG · Developer Tools · Knowledge SystemsSystem Design RAG Assistant
Large system-design and software-engineering books are difficult to search conversationally, while ungrounded LLM responses can introduce incorrect technical information.
PythonQdrantSentence TransformersLLMPDF Processing
Voice AI · NLP · Enterprise ProductivityAI MOM - Meeting Minutes Generator
Manually transcribing meetings and creating structured minutes, decisions, and action items is repetitive and time-consuming.
WhisperPythonFastAPILLMsNLP
NLP · Developer Tooling · Database IntelligenceAskQL - Natural Language to SQL
Business users often understand the question they want answered but lack the SQL knowledge required to retrieve the data themselves.
PythonspaCyFastAPIReactLLM IntegrationSQL
Speech AI · NLP · Production EvaluationASR Shootout - Indian Conversational Speech
Traditional ASR benchmarks based only on WER do not always reveal how models behave on noisy, conversational, multilingual speech or important named entities.
WhisperGoogle STTDeepgramJiWERPandas
MLOps · Fraud Detection · Production MLFraud Detection MLOps Platform
Fraud models require more than model training: reproducibility, experiment tracking, deployment, monitoring, and reliable inference are essential for production use.
PythonLightGBMXGBoostMLflowDockerFastAPI
MLOps · Production AI · Model MonitoringFake News Trend Drift Detector
A deployed model can silently degrade when production data changes, without an automatic mechanism to detect drift or trigger retraining.
FastAPIMLflowDVCEvidently AIPrefectGrafanaDocker
Generative AI · Synthetic Data · PrivacySynthetic Data Generation Platform
Real datasets can be difficult to share or use because of privacy constraints, while synthetic data must still preserve useful statistical relationships.
CTGANSDVPandasScikit-learnFastAPITypeScript
FinTech · Live Market SystemsAI-Powered Investment Portfolio Platform
Retail investors often lack integrated tools for live portfolio tracking, market analysis, alerts, and portfolio-level insights.
Next.jsPostgreSQLWebSocketsZerodha Kite ConnectJWT
Healthcare AI · Clinical Decision SupportAI Symptom Checker
Users often have difficulty interpreting symptoms or deciding what level of medical attention may be appropriate before consulting a professional.
PythonScikit-learnNLPRandom ForestGradient Boosting
Healthcare · Fraud Detection · MLHealth Insurance Fraud Detection System
Insurance fraud can involve subtle patterns across claims, providers, patients, and transaction attributes that are difficult to identify through simple rules.
PythonScikit-learnXGBoostPandasMachine Learning
Computer Vision · Infrastructure MonitoringStructural Defect Analyzer
Manual infrastructure inspection is expensive and time-consuming, particularly when cracks and other visible defects must be identified across large structures.
PythonTensorFlowKerasOpenCVCNNTransfer Learning
Financial Risk Analytics · ML · Anomaly DetectionCredit Card Fraud Detection System
Fraudulent transactions represent a very small percentage of transactions, making raw accuracy misleading for fraud detection.
XGBoostScikit-learnSMOTEPandasMatplotlib
NLP · Recommendation · Recruitment AIResume Skill Matcher
Recruiters and applicants need a consistent way to compare resume skills with job requirements without relying only on keyword matching.
PythonNLPTF-IDFSentence TransformersSimilarity Scoring
Data Engineering · Monitoring · AnalyticsData Quality & Pipeline Monitoring Dashboard
Data pipelines can produce incomplete, inconsistent, or structurally incorrect datasets without providing clear visibility into the health of incoming data.
PythonPandasSQLData Quality ChecksPower BI
