AI Engineer, Backend Engineer, and Automation Engineer
Backend, AI systems, and automation. Voice agents, multi-agent workflows, and manual processes turned into pipelines.
Voice agents held under 1000ms end to end.
Project work
CallCatalyst.ai
An AI call agent, end to end.
- Transport
- LiveKit handles real-time audio streaming over WebRTC; SIP trunking bridges to the PSTN so the agent answers ordinary phone numbers.
- Latency budget
- The full capture → STT → LLM → TTS → playback loop is held under 1000ms, the threshold where a caller stops experiencing the pause as a conversation gap.
- Orchestration
- LangChain manages turn-taking and conversation state across the call.
- Cost
- Operating cost held at 2.5–3 per minute of call time.
1000ms — the point a pause stops sounding like conversation
Stack: LiveKit · SIP trunking · JWT auth on session tokens · webhooks for call events · Redis · vobiz
Semantic caching: common questions are stored with their embedding as the key and the already-synthesised TTS audio as the value. A recognised question returns cached audio and skips speech-to-text, the model, and synthesis entirely.
Hover or tap a stage for its share of the relative duration. Per-stage split is representative; the 1000ms total is measured.
SmartInvest.ai
Company investment advisory: 20+ agents on a six-gate evaluation pipeline.
- Orchestration
- CrewAI agents handle multi-model orchestration, with each agent assigned a role in the evaluation pipeline: intake, six-gate qualification, comparable-company research, and IC memo generation.
- Retrieval
- A hybrid layer over Postgres (structured deal data), Neo4j (entity and relationship graph), and Qdrant (vector search over documents).
- Grounding
- Chunks are tagged by pipeline stage, so each agent retrieves only what is relevant to its gate and every output traces back to a real source document instead of inferred detail.
- Delivery
- FastAPI service, containerized with Docker.
20+ agents on a six-gate evaluation pipeline. Click a stage to trace its agents and the sources it touches.
CafePilot.tech
Order flow with a cache layer.
- Data model
- Relational Postgres schemas designed for concurrent transactions, so simultaneous orders do not collide.
- Caching
- A Redis layer in front of high-traffic read endpoints to cut repeated database load.
- Structure
- Microservices-oriented backend with a React front end and real-time order visualization.
- Result
- Order processing time cut 40%.
A cache hit skips the db hop.
What I build for clients
Automated calls that hold a real conversation. Real-time audio over WebRTC with SIP trunking so the agent answers an ordinary phone number, event-driven webhooks so your systems react to what happens on the call, and a latency budget tight enough that the caller never hears the machine thinking.
See CallCatalyst.aiA manual, repetitive process rebuilt as a pipeline that runs itself. Multi-agent orchestration where the work needs reasoning, n8n where it needs reliable plumbing between services.
See SmartInvest.aiRetrieval over your own files where answers trace back to a source document instead of being invented. Hybrid retrieval across relational, graph, and vector stores, with chunks tagged so each query hits the right slice.
See SmartInvest.aiFastAPI and Node services with JWT auth, Redis caching, Docker containers, and CI/CD, so what I hand over deploys the same way every time.
See CafePilot.techCapabilities
Building agent pipelines where the hard part is state and handoff between steps, not the prompting — CrewAI for role-assigned multi-agent orchestration, LangChain for conversational turn-taking, RAG where answers have to trace back to source documents.
FastAPI and NestJS services designed around the request patterns they actually serve, with a microservices split where it earns its complexity and a monolith where it doesn't.
Choosing the store to match the query — Postgres for relational and transactional work, Neo4j when relationships are the query, Qdrant for semantic search, Redis in front of hot reads.
Python for AI and backend services, JavaScript and TypeScript across the stack, Go and C/C++ from systems coursework.
Docker and GitHub Actions so services deploy the same way every time, unit tests, and Agile/Scrum sprint work at Codezilla.
Track
- Software developer
Build production REST APIs in FastAPI and NestJS, and integrate LLM-powered workflows that replace processes previously done by hand. Containerize services with Docker and maintain CI/CD pipelines for automated deployment. Work in Agile/Scrum sprints with unit testing and Git-based version control.
- Placement training
A year of aptitude and data structures and algorithms, alongside production-level project work.
- B.Tech, computer science engineering
CGPA 8.0/10.
- AWS Cloud Practitioner
- Oracle Academy Database Foundations
- Cisco Python Essentials 1 and 2
Contact
Start a project
In a first message: the problem, roughly how it works today, and any deadline.
- Phone
- +91 7049590456
- +91 7049590456
- LeetCode
- leetcode.com/u/adarshagrawal2004
Available for freelance and full-time work. Replies within a day.