Hiring used to take days. On Cannvue it takes under an hour. That gap, between how work is done and how it should be done, is where I spend my time.
| Focus | Detail |
|---|---|
| Building | Cannvue, AI interview platform. React + FastAPI, 16,300+ LOC across 49 files. Auth, real time interview flows, AI assisted scoring that replaces manual review. |
| Shipping | Intellious Technologies, production backend across 6 microservices and 219 endpoints. Design to deployment, 22 test files, CI/CD with GitHub Actions and Docker. |
| Researching | Fine tuning and RAG for domains where wrong answers cost money. Legal contracts, hiring decisions, business simulations. |
Cannvue, Founder (Mar 2026 to Present) : interviews that grade themselves
Full stack AI interview app in live use. 50+ users across 2 institutes. Weekly iterations driven by direct user feedback.
- Stack: React, FastAPI, PostgreSQL, WebSockets
- Scale: 16,300+ LOC, 49 source files
- Result: hiring cycles cut from days to under 1 hour
StartupVerse : a startup sandbox with 5 AI agents arguing with each other
Go backend, Python AI layer, React frontend. Simulates the full startup lifecycle with Narrator, Consumer, Competitor, Investor and Employee agents.
- Stack: Go, Python, React, PostgreSQL, Redis, WebSockets, Docker
- Scale: 59,200+ LOC across 450 files, 193 test files
- Core: deterministic business rules plus LLM driven agent behavior, RAG pipeline, real time state
Legal LLM work : fine tuning where accuracy is the product
- Fine tuned GPT-3.5 on 5,000 legal documents. LangChain plus Pinecone RAG pipeline at 85% accuracy on internal benchmarks.
- Fine tuned Llama 2-7B on 10,000 legal contracts with LoRA and QLoRA. 92% ROUGE-L on holdout, 500ms inference on EC2 g4dn.xlarge. Custom PDF extraction and semantic chunking.
Production backend, Intellious (Jun 2026 to Present)
Own services end to end across auth, user, timesheet and reporting. Integration and API suites that catch regressions before release. Code reviews and production debugging with senior engineers.
- Scale: 17,900+ LOC, 219 endpoints, 6 microservices
- Practice: tested releases, Docker deploys, Actions pipelines
1. Find work done by hand that repeats every week.
2. Build the smallest version that removes it.
3. Put it in front of users within days.
4. Measure. Keep what moves the number. Delete the rest.
Three rules I keep:
- Write the test that would have caught it. 193 test files on StartupVerse and 22 on Intellious services exist because regressions in hiring and payroll are expensive.
- Latency is a feature. 500ms inference did not happen by accident. Chunking, indexes and caching are product decisions.
- Ship on a cadence. Weekly at Azmth and Cannvue. Predictable releases beat large reveals.
- Pragati Engineering College, B.Tech Computer Science, 2023 to 2027. Coursework in DSA, databases, operating systems, networks, probability and statistics.
- Certifications: SAP ABAP Cloud Backend Developer, Salesforce Agentforce Specialist, Fast.ai Deep Learning Part 1, Microsoft Fabric DP-600 and DP-800, Google Cloud Generative AI.
- Freelance: LLM document Q&A and real time analytics dashboards. 5 star rating across Upwork contracts.
- CRM work at Azmth: reusable React and TypeScript components across internal apps. 20% faster page loads from lazy loading and code splitting.
Collaborations on applied AI, backend systems and dev tooling. If you are building something where correctness and latency both matter, I want to hear about it.
