RS
RISHI SINGH
Systems Directory

ENGINEERED SYSTEMS

A comprehensive catalog of applications, distributed pipelines, and ML experiments built around real engineering constraints, deterministic data validation, and verified codebases.

HEALTHCARE · AI · DATA QUALITYCompleted & Tested

VitalTrace

Personal Health Record Intelligence & Deterministic Data Quality

Full-stack clinical data intelligence system that ingests unstructured lab reports, executes deterministic physiological validation, normalizes test aliases to LOINC standards, and tracks longitudinal biomarkers with zero hallucination tolerance.

45 Vitest Specs Passing
  • ▸Engineered deterministic clinical validator engine enforcing biological plausibility bounds and unit standardization.
  • ▸Zero-hallucination LOINC code mapping: unrecognized tests retain raw string names with null codes.
React 19TypeScriptSupabasePostgreSQL RLSGemini APIZodVitest
DOCUMENT AI · HYBRID SEARCH · RAGCompleted & Tested

DocMind AI

Hybrid RAG Document Intelligence Platform with Page-Level Lineage

Enterprise document question-answering and synthesis engine combining FAISS dense vector search with BM25 sparse keyword matching via Reciprocal Rank Fusion, exact page citation pills, and dual Groq/Ollama inference backends.

Dense + Sparse RRF Engine
  • ▸Designed hybrid retrieval pipeline using Reciprocal Rank Fusion (RRF) to merge dense semantic and sparse lexical search.
  • ▸Engineered word-to-page boundary chunking guaranteeing zero-loss citation traceability to exact PDF pages.
PythonFastAPIFAISSBM25LangChainGroq CloudOllamaPandas
CLIMATE TECH · CONSTRAINT SCHEDULINGIn Development

WorkWise

Climate-Aware Construction Task Scheduling Engine

Predictive thermal strain and occupational safety platform for Indian construction operations, replacing crude clock rules with Liljegren (2008) physical WBGT modeling, CAQM GRAP regulatory constraint scheduling, and Bedrock AI briefings.

48/48 Pytest Passing
  • ▸Engineered physical heat-stress engine using peer-reviewed Liljegren (2008) formulation to calculate Wet-Bulb Globe Temperature.
  • ▸Strict architectural separation: deterministic Python safety engine decides constraint gates; Bedrock LLM only explains.
Python 3.13React 19AWS SAMLiljegren WBGTAmazon BedrockFastAPIPytest
DISTRIBUTED SYSTEMS · ASYNC PIPELINESCompleted & Tested

Dhara News

Distributed Indian News Verification & Bayesian Truth Engine

Distributed news intelligence platform with a 42-agent asynchronous RabbitMQ pipeline crawling 30+ outlets, featuring a Bayesian truth-scoring engine with N_eff source-independence correction and Kubernetes observability.

42-Agent Pipeline
  • ▸Engineered a distributed asynchronous pipeline with 42 worker agents communicating via RabbitMQ.
  • ▸Implemented a Bayesian truth-scoring engine with N_eff source-independence correction to prevent wire-story amplification.
PythonFastAPINext.jsRabbitMQPostgreSQLDockerKubernetesPrometheus
Technical SummaryGitHub
ASSISTIVE TECH · SEQUENTIAL MLCompleted & Tested

Voicefy (ISL Translation)

Real-Time Indian Sign Language Translation Platform

Assistive technology web application capturing sign language via live webcam video frames and translating sequential hand gestures into natural language speech using a Seq2Seq LSTM with attention.

Seq2Seq + Attention
  • ▸Trained a sequence-to-sequence LSTM with Bahdanau attention in TensorFlow/Keras on dynamic sign gestures.
  • ▸Engineered live video inference pipeline handling frame extraction, keypoint normalization, and padding in real time.
Next.jsReactFlaskTensorFlow/KerasSeq2Seq LSTMOpenCV
Technical SummaryGitHub
CLINICAL ML · SEQUENTIAL MODELINGResearch Prototype

Clinical Trajectory Modeling

Transformer-Based Sequential Health Record Prediction

Predictive healthcare modeling experiment using synthetic longitudinal EHR records (Synthea / MIMIC-IV paradigm) and transformer encoders to forecast next clinical condition onsets from historical diagnostic sequences.

Transformer Sequence Model
  • ▸Modeled longitudinal patient medical histories as discrete temporal sequences of diagnosis codes.
  • ▸Trained transformer encoder blocks with positional encodings to capture multi-year disease progression trajectories.
PyTorchPythonSynthea EHRTransformersScikit-learn
Technical SummaryGitHub