shivadeepak.dev
Back home
Project Catalog

Systems, workflows, and engineering artifacts.

A deeper look at the projects behind the portfolio, grouped by engineering signal rather than chronology.

Selected Work

Curated systems, not a project dump.

Public repositories are grounded in GitHub metadata and READMEs. Private or evolving systems are labeled as such.

Featured Systems

Retrieval, orchestration, telemetry, and document intelligence.

The strongest systems work: explicit routing, memory-aware reasoning, data pipelines, and agentic execution loops.

Agentic-RAG

● complete

Evaluation-driven retrieval and reasoning system over arXiv cs.AI papers.

A LangGraph-based agentic retrieval system that decides whether to retrieve, clarify, call a tool, refuse, or answer directly using memory-aware reasoning and multiple retrieval strategies.

System shape
User query
Decide
Memory context
Retrieve / tool / refuse
Grounded answer
  • Seven-node routing graph with retrieve, tool, clarify, refuse, answer, and chat paths.
  • Conversation, episodic, and semantic memory are injected into decision and answer prompts.
  • Vector, lightweight hybrid, true hybrid, and cross-encoder retrieval modes are benchmarked.
  • The evaluation covers retrieval, memory, refusal, clarification, tool routing, and smalltalk cases.
Engineering notes
  • Lightweight hybrid performed best on the benchmark; the cross-encoder added compute without improving retrieval quality.
  • Refusal and clarification are treated as first-class system outcomes, not error states.
PythonLangGraphFastAPIChromaBM25Sentence-TransformersGroqDocker

Deus Ex Machina

● active

Autonomous multi-agent orchestration framework for goal-driven execution.

An autonomous orchestration framework for goal-driven AI execution using LangGraph, vector-memory retrieval, and multi-agent planning loops.

System shape
Goal
Cognitive core
Router
Agent teams
Memory loop
  • Cognitive core decomposes goals into structured task plans.
  • Router dispatches work to specialist agent teams with distinct tool access.
  • Execution results are embedded into persistent vector memory for later runs.
  • Workspace isolation keeps file and terminal actions scoped to a sandboxed directory.
Engineering notes
  • The project explores long-running planning loops, not chat-only interaction.
  • Architecture is intentionally cyclic so outcomes can feed future planning steps.
PythonLangGraphGroq / Llama3ChromaDBOllama embeddingsPlaywrightRich

PromptLens

● active

Prompt telemetry, ETL, warehousing, and analytics platform.

An analytics and warehousing platform for LLM prompt telemetry with ETL pipelines, feature engineering, prompt observability, and behavior analysis workflows.

System shape
Raw logs
ETL
Feature engine
Warehouse
Analytics API
  • Raw prompt logs are normalized through schema adapters and validation layers.
  • Feature extraction maps prompt structure into warehouse-ready analytical fields.
  • PostgreSQL star schema, materialized views, and indexes support OLAP-style queries.
  • Analytics layer includes clustering, association rules, and baseline outcome models.
Engineering notes
  • The project frames prompt engineering as an observability and data problem, not a guessing loop.
  • Dashboard work sits on top of backend data workflows rather than replacing them.
PythonFastAPIPostgreSQLETLOLAPRNext.jsData Warehousing

Agent-Author

● private / activePrivate active system

AI document intelligence and publication pipeline.

A document intelligence pipeline combining OCR, translation workflows, semantic retrieval, commentary generation, export orchestration, and audit telemetry for structured publication systems.

System shape
Documents
OCR / translation
Semantic index
Commentary
Exports + audit
  • Treats document processing as a multi-stage workflow rather than a single generation step.
  • Combines extraction, retrieval, commentary, export, and audit concerns in one pipeline.
  • Architecture is private; public implementation notes are currently evolving.
Engineering notes
  • The useful part is orchestration: keeping provenance, generated commentary, and exports aligned.
  • Telemetry matters because document workflows need debuggable processing histories.
OCRSemantic RetrievalTranslationExport PipelinesTelemetryDocument Intelligence
Infrastructure & Automation

Daemon systems, event loops, media pipelines, and long-running agents.

Backend-heavy projects that coordinate background work, telemetry, assets, and external tools.

Senthium-AI

● active

Presence-aware workstation monitoring and automation platform.

A daemon-based workstation monitoring and automation platform with telemetry collection, rule-based triggers, dashboard orchestration, and real-time event monitoring.

System shape
Daemon
Rule triggers
Telemetry log
Dashboard
Alerts
  • Background daemon and service commands support long-running workstation monitoring.
  • Rules engine evaluates process, CPU, network, disk, and schedule triggers.
  • Activity logging, analytics summaries, and multi-channel alerting make behavior observable.
PythonOpenCVDeepFaceStreamlitIPCJSONL telemetryCLIWebSocket

project-storybook

● active

AI-assisted multimedia orchestration pipeline.

An automation pipeline that coordinates story generation, asset extraction, and FFmpeg-based synthesis workflows for narrated multimedia content.

System shape
Source
Planning
Asset extraction
TTS + subtitles
FFmpeg render
  • FastAPI backend manages background generation work and streams stage updates over SSE.
  • Playwright extraction aligns generated text and image assets into a structured story file.
  • FFmpeg filtergraphs, SSML, subtitles, and scene stitching produce the final media artifact.
Engineering notes
  • The engineering value is in durable media orchestration and fallback handling, not content novelty.
PythonFastAPINext.jsSSEPlaywrightFFmpegAzure SpeechNVIDIA NIM

project-LIFE

● experimentalNo public repository yet

Autonomous Discord-based AI workflow agent.

An asynchronous AI agent with persistent memory, intent-driven execution, multimodal generation, and long-running workflow orchestration within Discord environments.

System shape
Discord event
Intent router
Memory
Tools / media
Long-running job
  • Architecture is experimental and currently evolving.
  • Focus is on persistent memory, async event handling, and workflow continuity.
  • Presented as a research-oriented agent system rather than a finished product.
Async agentsDiscord workflowsPersistent memoryIntent routingMultimodal generationAutomation
Foundations

Earlier full-stack systems work.

Foundational engineering work that shows comfort with persistent state, role-based workflows, and larger application surfaces.

bank-management-system

● foundational

Large foundational full-stack banking workflow system.

A full-stack banking application with customer and admin workflows, transactional operations, loan processing, and persistent account management built using Node.js, Express, EJS, and MySQL.

System shape
Browser
Express routes
Handlers
MySQL
Admin / customer workflows
  • Customer and admin portals cover account opening, approval, transactions, loans, and profile workflows.
  • MySQL-backed request flow uses Express routes, session authentication, and server-rendered views.
  • Positioned as earlier systems experience, not as a current AI infrastructure project.
Node.jsExpressEJSMySQLbcryptexpress-sessionjQuery