Building AI products at SIP Organization · Former AI Engineer Intern at micro1
I build the parts an impressive demo can hide: retrieval that exposes its evidence, agents that respect permission, and services that recover when things go wrong.
Enter the portfolio → · 60-second recruiter view · Résumé
LinkedIn · Email · X / Twitter
Bengaluru, India · Open to AI engineering, full-stack roles, internships, and contract builds
Autonomous Personal Agent — control before autonomy
Local alpha · Python / FastAPI / PostgreSQL / Redis / Docker
Built a self-hosted agent control plane where task, outbox, and audit state share a transaction. Worker leases, heartbeats, bounded retries, and dead letters make interrupted work recoverable. Exact-action approvals bind reviewed content to a digest; persisted receipts prevent replay of recorded side effects.
The decision: refuse a stale worker's completion instead of trusting that only one process is still running.
Lifecycle decision · Approval + receipt implementation · CI evidence
Current boundary
Private local dashboard, career workflows, and isolated action adapters. External submissions and email remain exact-action approval-gated; broad browser autonomy is not enabled. Local fixture verification is not a claim of universal ATS compatibility. Exact-action design.
AtlasLM — make the answer inspectable
Live · Next.js / TypeScript / Qdrant / Upstash Vector / OpenRouter
Built a document-intelligence workbench with deterministic ingestion, dense + BM25 retrieval, reciprocal-rank fusion, reranking, and MMR diversity. The answer arrives with source passages, citation checks, and timed execution traces—not just generated prose.
The decision: gate before generation and audit after it. Weak evidence can produce abstention instead of confident-looking text.
Try the live workspace · Inspect the pipeline · Quality gates
Current boundary
One source workspace at a time. The semantic cache is instance-local; a multi-instance service would need shared cache state. Qdrant serves the local path and Upstash Vector the serverless deployment. Architecture and honest scope.
MinePulse / KarixMC — a real game-to-web boundary
Live on VPS · Next.js / Java / Paper / PostgreSQL / Redis / Nginx
Built and deployed a Minecraft marketplace connecting verified playtime, reward ledgers, server stores, and a Paper plugin. Signed events cross the game/API boundary; expiring purchase claims and a plugin-side receipt journal handle delivery acknowledgements.
The decision: calculate rewards on the server and remember delivered commands, so a lost acknowledgement need not repeat a recorded delivery.
Explore KarixMC · Inspect the plugin · Security boundary · Production runbook
Current boundary
Two application replicas behind Nginx, PostgreSQL, Redis, and recurring encrypted backups. Campaign-credit purchases are manually confirmed; automated payment checkout is not connected. Deployment is not a claim of customer scale. Current implementation.
- ROLLFORWARD — optimistic UI under failure: separate projected state from confirmed truth, preserve intent in IndexedDB, retry with the same idempotency key, and make
412conflicts explicit. Source + decisions - Revive — live payment-recovery prototype with deterministic policies, HMAC-verified webhooks, and immutable audit records. Its duplicate-suppression challenge exercises the hosted backend; payment outcomes are demo data, not real charges or measured merchant uplift. Source
- AtlasForge AI — checkpointed video production with provider fallbacks, editorial gates, and bounded cloud spend. SQLite and filesystem artifacts keep the single-workstation system deliberately simple. Architecture
Lower-level explorations
- GPT Prototype: a decoder-only transformer built from first principles with RoPE, RMSNorm, QK-Norm, and Muon + AdamW.
- Distributed Search Typeahead: FastAPI, PostgreSQL, Redis Streams, consistent hashing, batched writes, and failure handling.
- SIP Organization · Project Lead Developer Intern · Jul 2026–Present — AI-powered WhatsApp onboarding, conversational workflows, backend integrations, and service delivery.
- micro1 · AI Engineer Intern · Aug 2025–Jul 2026 — worker orchestration and modular cell components for a self-adaptive AI architecture.
- Independent AI / ML Developer · 2025–Present — applied LLM products, automation, APIs, and training/inference workflows.
B.Sc. Computer Science, Scaler School of Technology with BITS Pilani · Aug 2024–Sep 2028. micro1 AI/ML credential
Working stack
AI: LLM applications, agent workflows, RAG, hybrid retrieval, evaluation, citations, PyTorch, OpenRouter.
Application: Python, TypeScript, FastAPI, Next.js, React, Java, REST APIs, OAuth, signed webhooks.
Data + delivery: PostgreSQL, pgvector, Redis, Qdrant, Upstash Vector, Prisma, Docker, Linux, Nginx, GitHub Actions, Vercel, VPS.
Public build activity
Generated daily from public GitHub data. Activity is context, not a substitute for the source and engineering decisions above.
Have a system worth building? Tell me the problem.
