SmaranQ
Private, local-first macOS meeting intelligence in alpha: local transcription, screen context, retrieval and a compact in-meeting HUD (Q) that surfaces timely cues.
Context Engineer / Senior AI Developer at Accellor / IIT Mandi
I turn context and agent workflows into working software:
from enterprise AI copilots to macOS tools and open-source projects such as SmaranQ, Camoscope, ActionNova and Agentarium.
Senior AI Developer at Accellor, building stateful AI agents, enterprise RAG, and reliable multi-agent workflows.
Current focus: Context + RAG + Agents
Mode: Production AI + native products
Writing: Claude Code Field Guide
About
I build software that can retrieve the right context, use tools, retain state and recover when workflows fail. At Accellor, my work covers production AI copilots, Microsoft Agent Framework, Azure OpenAI and Azure AI services, FastAPI/SSE, and human-in-the-loop orchestration.
Before agentic AI, I built revenue-intelligence models at RevSure AI, ML targeting and recommendation systems at Pixis, and product analytics at LEAD School. I earned a Civil Engineering B.Tech with a minor in Management from IIT Mandi after studying at Abhayanand Super 30.
My independent work includes SmaranQ (meeting intelligence), Camoscope (consent-based screen-share auditing), ActionNova (macOS utility workspace), and the open-source Agentarium project. OrgSmith is still an architecture concept and HTML prototype, not a functioning agent-execution engine.
Selected systems
Products, experiments and agent systems
Featured work first, followed by early prototypes and experiments. Release status is shown on each project.
Camoscope
Camoscope v0.1.3 is a consent-based macOS screen-share audit CLI (public beta). It flags windows potentially excluded from capture, identifies owning apps and can inspect Accessibility-exposed text; findings require verification against the actual share.
ActionNova
ActionNova v0.1.14 is a native Swift 6.2/macOS 26+ development build that turns the notch into an ambient workspace for timers, media, camera, calendar, clipboard, files, notes, Shortcuts and system activity.
Agentarium
Agentarium 0.3.0 is a local-first browser workspace for human-led AI-agent teams: connect a project folder, edit tasks, switch project phases and save/export locally. Maestro is a simulation preview; real agent execution is upcoming.
Production systems
Professional AI work
Public-safe architecture summaries from current production work. Client-specific details remain private.
Production AI copilots & multi-agent systems
Copilot workflows with persistent agent sessions, structured outputs, retries, checkpoints, state transitions and human-in-the-loop approvals.
Context and institutional memory
Retrieval and memory layers spanning enterprise documents and prior interaction context using Azure AI services, search, Redis, and persistent state.
AI application infrastructure
FastAPI/SSE services around Azure OpenAI, AI Foundry, Document Intelligence, search, storage, and production agent workflows.
More projects and experiments
OrgSmith
Concept and HTML prototypes for a hierarchical AI-organization simulator with specialist-to-founder escalation, token-budgeted shifts, human oversight and artifact-first workflows. The orchestration engine is planned, not yet implemented.
SplitKar Platform
Expense-splitting web app with a Next.js/React frontend and Node.js/Express/PostgreSQL backend, including group expenses, settlements, authentication and role-based access controls.
Agentic Data Analysis
Python CLI experiment using a CrewAI-inspired multi-agent workflow to clean, analyze and answer natural-language questions about a flight-booking dataset.
Recommendation System
Recommendation-system experiment comparing confidence-based and bias-vector ranking models using Precision@10 and NDCG@10, with evaluation results documented in the repository.
Sparkpad
Local-first browser Markdown editor with live preview, GitHub-flavored Markdown, syntax highlighting, KaTeX maths, Mermaid diagrams, autosave and .md export.
Timer / Clock / Stopwatch
Real-time timekeeping utility app shipped end-to-end.
ZenConnect
Early-stage chat and collaboration interface concept associated with SplitKar, developed as a separate UI prototype.
Sheetwise Expense Buddy
Utility app for streamlined expense workflows and summaries.
Invitation / Lifestyle Sites
Minimal single-purpose sites for invitations, plans, and quick static experiences.
Experience
Senior AI Developer, Accellor
Production AI copilots & multi-agent systems: Copilot workflows with persistent agent sessions, structured outputs, retries, checkpoints, state transitions and human-in-the-loop approvals.
Context and institutional memory: Retrieval and memory layers spanning enterprise documents and prior interaction context using Azure AI services, search, Redis, and persistent state.
AI application infrastructure: FastAPI/SSE services around Azure OpenAI, AI Foundry, Document Intelligence, search, storage, and production agent workflows.
Data Scientist, RevSure AI
Lead Fit scoring system: Prioritized accounts using firmographic, behavioral, and conversion signals to improve sales targeting efficiency.
Lead Propensity model: Predicted progression across the funnel (MQL→SQL→Opportunity→Closed-Won) to improve forecasting and campaign optimization.
Title Standardization: LLM + rule-based cleaning/normalization for CRM job title data to improve downstream model performance and reporting.
Probabilistic multi-touch attribution: HMM-based attribution to quantify channel impact and move beyond static rule-based attribution.
Buying Stage inference: Combined product, marketing, and sales signals to infer customer journey stages and provide real-time visibility.
LLM-based data standardization: Applied LLM-based semantic normalization techniques to reduce manual effort and improve data quality across pipelines.
Data Scientist, Pixis
Predictive targeting & optimization models: Behavioral, engagement, and contextual signal models to improve campaign performance.
Marketing data pipelines: Built real-time + batch processing pipelines using AWS S3 + Athena for ML inference and reporting.
Feature engineering framework: Reusable feature creation workflow to transform raw signals into ML-ready inputs.
Codeless AI infrastructure: Enabled non-technical users to leverage ML insights in workflows.
Early generative AI experiments: Prompt-based experimentation and integrations for targeting/segmentation automation.
Product Analyst I, LEAD School
Product dashboards & reporting: Built interactive dashboards to track performance and business metrics for product teams.
Reporting pipeline automation: Automated reporting workflows using Python + Airflow, reducing manual reporting effort.
Student Representative, Career and Placement Cell, IIT Mandi
Placement policy & stakeholder communication: Improved student placement experience and policy execution through coordination with academic departments and administration.
Intern, Geo Carte Radar Technology Pvt. Ltd.
Crack detection POCs: Led POCs and problem-solving for crack detection using tooling and data interpretation.
Data quality & troubleshooting: Troubleshot technical issues to ensure accurate submissions and data integrity for contractors.
Skills
AI Engineering
AI Platform & Infrastructure
Software & Data
Writing
Claude Code for Beginners: 10 Habits That Change Everything
Ten habits that make Claude Code feel less random and more like a useful coding partner.
Claude Code Advanced Workflow: The Full Stack
A full-stack workflow for planning, model choice, memory, review, hooks, skills, and parallel execution.
Hooks: Give Claude Code Guard Rails That Actually Run
How to turn repeated reminders into Claude Code hooks that actually run.
Skills: Repeatable Process On Demand
Why reusable skills beat ad-hoc prompts for engineering workflows that need process.
Slash Commands: The Power Shortcuts Most Developers Overlook
A practical map of the slash commands worth building into Claude Code muscle memory.
Plan Mode: Think Before You Code
How plan mode separates deciding what should change from writing the code.
Commands vs Skills: Know the Difference, Use Both Right
A modern decision rule for choosing between prompt-like commands and structured skills.
More writing
Notes on developer tooling, data science, and practical workflows.
How to Show Inline Function Arguments Using Extensions and Settings in VS Code
A practical guide to Python inlay hints in VS Code: configure Pylance to display function parameter names where you need them.
Google Marketing Live 2024: AI Revolutionises Advertising
A look at the AI-driven advertising and creative features announced at Google Marketing Live 2024.
VSCode Mastery for Data Scientists: Essential Extensions and Productivity Hacks
A curated guide to VS Code extensions and small workflow improvements for day-to-day data science.
Resume
Formal wording: Senior AI Developer at Accellor, Data Scientist at RevSure AI and Pixis, Product Analyst I at LEAD School.
Life
Early years with my maternal grandparents
The first chapter was rooted in Chauri, around my maternal grandparents' home. It gave me a grounded start: family, discipline, and the kind of curiosity that comes from watching people solve everyday problems with limited resources.
- Grew up close to family and village life
- Built early curiosity through observation
- Carried a grounded sense of effort into school
Village-school foundation through Class 2
Primary schooling started in my village, Kanhaiyachak. It was a simple beginning, but it built the first habits of showing up, learning steadily, and staying close to the realities around me.
- Early schooling through Class 2
- Village-first learning environment
- Built consistency before moving to Jehanabad
Middle-school years at Shanti Kunj Public School
At Shanti Kunj Public School, the academic world widened. This was where school became more structured and I started building confidence across subjects.
- Moved from village schooling into Jehanabad
- Built stronger academic rhythm
- Continued developing interest in science and problem solving
Maths and physics foundation at Bal Vidya Niketan
Bal Vidya Niketan became the phase where maths and physics started standing out. The focus shifted from just doing well in class to preparing for a larger academic path.
- Strengthened mathematics and physics
- Built exam discipline and competitive intent
- Prepared the base for science and JEE preparation
Science and JEE preparation at Abhayanand Super 30
The Super 30 phase in Patna sharpened everything: science, JEE preparation, independence, and the belief that a student from a small-town background could compete at a national level.
- Studied science with a strong maths and physics focus
- Prepared intensively for JEE
- Built resilience, discipline, and independent thinking
B.Tech in Civil Engineering with a management minor
IIT Mandi gave me a technical foundation through Civil Engineering, plus a management minor and a wider world of projects, leadership, and problem-solving. It was also the bridge from academic strength into product and data thinking.
- B.Tech in Civil Engineering with a minor in Management
- Built technical, analytical, and leadership foundations
- Represented IIT Mandi in technical competitions and campus initiatives
From analytics and data science to agentic AI engineering
The professional chapter started with product analytics, moved into production data science and revenue intelligence, and now sits in agentic AI systems: RAG, multi-agent workflows, AI agents, and generative AI products.
- Product analytics and reporting automation at LEAD School
- Production ML, campaign optimization, and data systems at Pixis and RevSure AI
- Current focus: agentic AI systems and multi-agent workflows at Accellor
Contact
Preferred channel: contact@rajsharma.space.