One person who ships and runs production AI systems. RAG, vector search, real-time pipelines, built, deployed, and maintained. No team behind the repos; it's all verifiable.
📧 parbhat@parbhat.dev · 🔗 parbhat.dev · 💼 Open to full-time remote (US/EU startups)
- Live products you can use today: Sentinel, Visura, VectorMail. All with auth, billing, and production practices.
- Measurable impact: 50-80% AI cost reduction (Visura), sub-250ms risk scoring (Sentinel), semantic search across 10K+ emails (VectorMail). Not claims - implemented in the repos above.
- Full-stack ownership: Schema → API → frontend → monitoring. I've run these systems myself for years; I fix root causes, not symptoms.
- Looking for: I want to own systems entirely and ship without hand-holding.
If you need someone who can take "we need RAG / pipelines / internal AI" from zero to production and keep it running, I've done that multiple times. Open to a conversation.
Most candidates have either: shipped production code but not AI/ML, or done AI demos that never went to production. I've done both - multiple production RAG systems, real-time data pipelines, and cost-optimized LLM infra, all live in production and maintained by me.
What you get:
- Ship speed. I don't need a long ramp. I've built and operated similar systems; I know the failure modes and how to avoid them.
- Full ownership. Give me the problem and the constraints; I'll own design, implementation, deployment, and ops. I document and hand off cleanly when the team grows.
- Cost-aware AI. I've cut per-document processing from $5 to $0.05 via architecture (Visura), reduced redundant LLM calls with caching and deduplication, and used multi-provider routing so you're not locked to one vendor.
- Production discipline. Health checks, structured logging, retries, graceful degradation. When something breaks, I debug it and fix the cause, not the symptom.
| What | Evidence |
|---|---|
| Live products | Sentinel, Visura, VectorMail. Auth, integrations, billing where relevant. All maintained. |
| Impact metrics | 50-80% AI cost reduction via hash-based chunk reuse (Visura); sub-250ms risk scoring (Sentinel); semantic search across 10K+ emails (VectorMail). See project sections. |
| Stack | Next.js 15, TypeScript, PostgreSQL, pgvector, Redis, OpenRouter, Sentry, OpenTelemetry, Clerk, Stripe. Repos and package.json are public. |
| Production practices | Self-healing recovery, cost guardrails, full observability (Sentry + OTel), distributed rate limiting. Focus on debuggability and automatic recovery. |
| GitHub | github.com/parbhatkapila4 - Sentinel, Visura, VectorMail (and more). Active in RAG, vector search, and full-stack TypeScript. |
These are not weekend projects. Each has the kind of structure (env, errors, docs, monitoring) that shows I can ship and maintain systems at startup pace.
Problem: CRMs show status, not risk. Deals stall and you find out too late.
What I built: Live pipeline ingestion (webhooks + sync), time-decay and stage-velocity modeling, explainable risk scores so teams see why a deal is flagged. Sub-250ms, real-time.
Why it's serious: Idempotent webhooks, dedup, retry/backoff, queue-based processing so bursts don't kill the API. Health checks and graceful degradation when CRM/calendar APIs fail. Live · Code
Problem: Reprocessing documents wastes AI budget. No visibility into costs, no recovery when things fail mid-process.
What I built: Enterprise-scale document intelligence platform. Hash-based chunk reuse saves 50-80% on AI costs. Self-healing processing with automatic crash recovery, full observability (Sentry + OpenTelemetry), distributed rate limiting, and cost guardrails. Vector search with 85%+ embedding cache hit rate. P50 under 2.5s for document processing.
Why it's serious: Versioned processing with idempotent replay, multi-layer security (HMAC signing, Clerk JWT, Zod validation), workspace collaboration with RBAC, and comprehensive monitoring. Not a demo - production architecture with real cost controls. Live · Code
RAG in production - Multiple production systems, different chunking and retrieval strategies. Vector search (pgvector, HNSW), persistent embeddings, context-only synthesis to control hallucination. See Visura and VectorMail.
Cost-efficient AI - Chunk dedup (Visura), Redis and in-memory caching, OpenRouter for model routing. Track and minimize cost per query where it matters.
Production ops - Health checks, structured logs, retry/backoff, graceful degradation. Deployed on Vercel and similar; I've run and debugged these systems myself.
Real-time and scale - Idempotent webhooks, queues, background jobs, indexing at scale. Time-series and vector indexing, connection pooling, latency-aware design.
Full-stack - Auth (OAuth), billing (Stripe/PayPal/..), RESTful APIs, clear error handling. Entire ownership from DB to UI.
- Async-first. I document decisions, write clear PRs (why, not just what), and don't need hand-holding.
- Ownership. I ship and maintain. I think about monitoring, cost, and docs because I've had to operate what I build.
- Debugging. I go for root cause, not workarounds. I've fixed production issues and built the practices so the next one is easier to find.
- What I need: Clear problem and constraints, feedback (code/design review), and trust to own my scope Entirely
I'm not listing buzzwords. Here are areas I can go deep on - and have already implemented:
RAG: Chunking strategies, hallucination control, citation tracking, embedding vs cost tradeoffs.
Production: Debugging slow queries, webhook retries without duplicates, monitoring and alerting, DB performance under load.
Cost: Per-query LLM cost, caching and dedup, when to use GPT-4 vs Claude vs Gemini.
System design: Rate limiting, error handling, reliability when dependencies are outside your control.
The repos above are the evidence. Happy to walk through any of it live.
I want: Seed to Series A, US/EU. Founding or early engineer role where I own systems. Small teams (<20), high ownership. RAG, AI tooling, data pipelines, internal tools, SaaS infra.
I don't want: Agencies, ticket-only roles with no design input, or teams with no technical leadership to learn from.
How to reach me:
Email: parbhat@parbhat.dev (best)
LinkedIn: linkedin.com/in/parbhat-kapila
I reply within 24 hours to serious outreach.
Good first email: Problem you're solving, who I'd work with, what I'd own, and growth path. I read every thoughtful message and reply even when it's not a fit.
Live: parbhat.dev
Source for the portfolio: SEO, link previews, fast load.
npm install && npm run devOpen http://localhost:3000. Production: npm run build → npm start.
Stack: Next.js 16 (App Router), React 19, TypeScript, Tailwind 4, Motion, next-themes, PostHog, Vercel. Meta/OG, sitemap, robots, accessibility (skip link, focus, reduced motion).
Private. All rights reserved.