United States of America
Coming soon...Travel photo
Sankalp's Portfolio Assistant
Online · trained on resume & projects
Replies use a Vercel serverless function ·
GROQ_API_KEY stays server-sidePlaces I've travelled
A clean travel shelf for the countries I have visited, with space to drop in one favorite photo from each place.
8Countries
1Photo Each
India
Coming soon...Travel photo
France
Coming soon...Travel photo
Italy
Coming soon...Travel photo
Switzerland
Coming soon...Travel photo
Germany
Coming soon...Travel photo
Netherlands
Coming soon...Travel photo
Austria
Coming soon...Travel photo
Travel photos coming soon.
Things I've built & shipped
Production systems, applied research, and the kind of work that ships on Friday and stays up on Monday. Tap any card to flip →
1.5+
Years
220K+
Embeddings
3.81
GPA
Experience
Stealth AI Startup
Software Developer Intern · Backend / AI · Remote
Designed and shipped REST API endpoints for user onboarding and profile standardization in a consumer app for physical-trade job matching. Owned backend development across all application components for the startup's physical AI product.
Flip
What I'm doingStealth Physical AI Startup · Backend
- Designed and shipped REST API endpoints for user onboarding and profile standardization in a consumer app for physical-trade job matching.
- Owned backend development across application components for the startup's physical AI product, including API contracts, service logic, data flow, and integrations.
- Built a semantic role classification pipeline with pgvector and sentence embeddings mapping titles to ~900 O*NET occupations - 95% accuracy in offline eval.
- Chose pgvector over a standalone vector DB to reduce ops complexity and keep embeddings colocated with profile data.
- Wrote PyTest unit + integration tests and added structured logging; aligned roadmap weekly with EM and founder.
- Collaborated with a 2-person frontend team to define API contracts and ship against weekly milestones.
Quinnox Inc. / Qyrus
AI Engineer Intern
Built a production RAG system indexing 220K+ embeddings. Improved precision@5 by 22% via CLIP + Cohere ReRank v3. Fault-tolerant ingestion at 20K+ items/hour on AWS.
Flip
What I shippedQyrus RAG Platform
- Built and deployed a production RAG system (FastAPI, Redis, Pinecone, Docker) indexing 220K+ embeddings for low-latency search.
- Designed embedding + reranking workflows with CLIP and Cohere ReRank v3 - +22% precision@5 while balancing latency and cost.
- Architected a fault-tolerant async pipeline on AWS (S3, SQS, async workers, batch upserts) with backpressure and retry logic - 20K+ items/hour.
- Sustained 300+ vectors/min ingestion throughput with structured metrics and distributed logs.
- Integrated LLM generation with grounding, prompt templates, and fallback logic; validated end-to-end with PyTest + Postman.
Pcloudy / Opkey
Software Engineering Intern
Shipped Qpilot.AI MVP - automated test generation across 1,200+ OEM configs with 78% execution success and 90% reduction in test time. Kafka log pipeline at 33K+ events/min.
Flip
What I shippedQpilot.AI MVP
- Designed and shipped the Qpilot.AI MVP (Flask, Python) - automated script generation and self-healing execution across 1,200+ OEM device configs.
- Achieved 78% execution success and reduced end-to-end test time by 90%, significantly lowering manual intervention.
- Built Python/Flask microservices on AWS with Kafka-based streaming log pipelines for real-time failure triage at 33K+ events/min.
- Automated build, test, deploy via GitHub Actions; containerized with Docker for iterative rollouts.
- Validated services with Postman + PyTest; added logging and monitoring for production reliability.
Projects
Sage
Multi-agent orchestrator & personal AI assistant
Local-first multi-agent system with RAG, persistent memory, live news, web search, WhatsApp delivery, Gmail triage, and proactive briefings - orchestrating specialized agents via Ollama tool-calling. No cloud, no tracking.
Flip
Project deep-diveSage · Multi-Agent Orchestrator
- Multi-agent orchestration via open-source Ollama tool-calling - a router LLM autonomously dispatches to specialized agents (web search, news, RAG, memory, reminders, Gmail) based on intent.
- Local-first RAG pipeline over personal documents (PDF, Markdown, TXT) with ChromaDB vector search and selective retrieval - skips embedding overhead for casual chat.
- WhatsApp integration via Twilio + FastAPI webhook: full chat, commands, and proactive reminder delivery with daily scheduling and missed-reminder recovery.
- Gmail triage agent via Google OAuth - classifies inbox into ACTION / FYI / IGNORE with AI-generated summaries surfacing what needs a reply.
- Persistent SQLite-backed sessions, facts, and todos with session resume, conversation analytics dashboard, and usage tracking.
Fandom Wiki Search Engine
Search engine for fan-driven knowledge bases
Final project for CS242: Information Retrieval and Web Search. Indexes and retrieves relevant content across Fandom wikis so users can quickly find characters, lore, game mechanics, and community knowledge.
Flip
Project deep-diveFandom Wiki Search Engine
- Built a specialized search engine across Fandom wikis for CS242: Information Retrieval and Web Search.
- Indexed fan-driven knowledge bases so users can retrieve relevant information on characters, lore, game mechanics, and community content.
- Implemented keyword-based retrieval and ranking techniques to return accurate, fast results for fandom-specific queries.
- Designed around the needs of fandom communities, where search intent often depends on names, aliases, game terms, and lore-specific context.
SafeBot
Investigation platform for Riverside PD
Human-in-the-loop email triage and suspect-flagging with hybrid keyword rules + LLM risk scoring. Google OAuth, RBAC, audit logs. Deployed on GCP Cloud Run with CI/CD.
Flip
Project deep-diveSafeBot · Applied AI for Investigations
- Built a production investigation platform for Riverside Police Department processing 10K+ emails/hour for triage and suspect flagging.
- Hybrid keyword rules + LLM-based risk scoring with RBAC, escalation policies, and review gates aligned to legal requirements.
- Shipped REST APIs with PostgreSQL schema migrations and React + TypeScript dashboards with Google OAuth, admin console, and audit logs.
- Deployed on GCP Cloud Run with Redis caching and autoscaling via GitHub Actions CI/CD.
- Currently in pilot discussions for deployment to 5 RPD investigators.
Android Log Search Utility
Real-time log parsing & classification
Parsed 250K+ Android log lines with failure classification. Published in the International Journal of Applied Engineering and Technology, London, 2024.
Flip
Project deep-diveAndroid Log Search Utility
- Built a real-time Android log parsing and search system (Node.js, MongoDB) processing 250K+ log lines.
- Added failure classification and recommendations for log lines, surfacing error patterns from noisy logs.
- Improved debugging efficiency for engineers working across high-volume device fleets.
- Work published as "Application Specific Log Parser" in the International Journal of Applied Engineering and Technology, London, 2024.
Test Device Recommendation Engine
ML-driven device matrix selection
Clustered 1K+ OEM configurations and ranked similarity over 125K+ log lines via TF-IDF. Replaced manual heuristics with data-driven scoring and cut device selection time to under 90 seconds.
Flip
Project deep-diveTest Device Recommendation Engine
- Built a recommendation system to automate test-device selection for QA teams managing massive device matrices.
- Clustered 1K+ OEM configurations and analyzed 125K+ log lines using TF-IDF and similarity-based ranking.
- Replaced manual heuristics with data-driven scoring - device selection time cut to under 90 seconds.
- Integrated as a recommendation layer into the broader Pcloudy test automation platform.
Education
University of California, Riverside
MS in Computer Science
GPA
3.81 / 4.0
PES University, Bangalore
BTech CS · Machine Intelligence & Data Science
GPA
8.3 / 10
Publications
Application Specific Log Parser
International Journal of Applied Engineering and Technology
Real-time Android log parsing, failure classification, and search across 250K+ log lines - improving debugging efficiency for engineers working across high-volume device fleets.
View Paper