Hi, I'm Rahul Bhardwaj 👋

Backend & AI engineer (fintech) — I build systems that move money and tools developers actually download.

---

5 years shipping Node.js / NestJS microservices in regulated fintech. My work has powered a direct credit-bureau integration that saves ₹30L+ annually, helped enable ₹70Cr+ in projected co-lending disbursements, migrated 20+ microservices to EKS (~40% infra cost reduction), and taken a critical billing job from 6 minutes to 11.2 seconds.

Now building at the intersection of backend systems and applied AI — agent orchestration, local-first voice AI, and developer tooling in the open. My Claude Code analyzer inspecto has crossed 3,000+ downloads on npm.

---

🛠 Stack

!Node.js
!NestJS
!TypeScript
!Python
!PostgreSQL
!MongoDB
!Redis
!RabbitMQ
!AWS
!Kubernetes
!Docker

> Daily drivers: NestJS · TypeScript · PostgreSQL (+ pgvector) · MongoDB · Redis · RabbitMQ · AWS (SQS · S3 · EKS · Lambda) · Kubernetes · Docker · Claude API / MCP

---

🔧 Open-source & AI projects

*Ranked by real-world impact — downloads, users, and problems solved.*

1. 📦 inspecto

Claude Code session quality analyzer — 12 behavioral & cost metrics pulled straight from your session logs.

!npm
!downloads

3,000+ npm downloads

2. 🛡 veto

Cost governor for Claude Code — a plugin that enforces spend policies before your agent burns your budget. Policy rules, session budgets, hard stops.

220+ tests passing

3. 🕸 agentrie

NestJS multi-agent orchestration & observability platform — distributed agent coordination with W3C trace propagation and a full eval/scoring layer: dataset runner, trace-derived scorers, LLM-as-judge, and compare mode with regression detection.

88+ tests passing · CLI exits non-zero on regressions

4. 🎙 Vani (वाणी) वाणी

Fully local, offline voice dictation for macOS (Apple Silicon) — WhisperKit + CoreML, zero network, zero data leaving your machine. Vani launched publicly; Uvaach is its open-source evolution, positioned as a Wispr Flow alternative.

On-device ML · Privacy-first · Launched

5. 📝 gradelee

AI homework grader — snap a photo of handwritten work → OCR → RAG retrieves syllabus context → adaptive feedback + grade. Serverless on AWS (API Gateway · Lambda · CDK), end-to-end in ~8s.

Model sits behind one swappable provider seam (env-var driven) — hosted fast model in the sync path, and flips to fully offline Gemma 12B for zero-network validation. Eval harness scores grounding, accuracy, and p50/p95 latency across models.

Built AI-augmented with Claude Code in 4 days

6. 🔍 traceglass

Tamper-evident audit logging for AI agents — SHA-256 hash-chained logs so you can prove what your agent did (and didn't do). Built for the coming wave of AI compliance requirements.

Hash-chain integrity verification

7. 🧠 memory-os

Local-first AI memory vault — NestJS + PostgreSQL + pgvector, exposed over the Model Context Protocol. Long-term memory for any MCP-capable agent.

Active

8. ⏰ timebomb

Static analysis for code that fails later — setTimeout int32 overflows, unbounded Promise.all, sequential awaits in loops.

🔥 Found real bugs in inspecto itself

9. 📊 trendrie

YouTube intelligence platform for tracking AI/dev-tools content trends — what's rising, what's saturated, where the gaps are.

In development

10. 🎤 Local mock interview coach

Fully offline interview practice app — Electron + Python sidecar, mlx-whisper for transcription, MediaPipe for delivery analysis. Practice out loud, get feedback, nothing leaves your laptop.

On-device AI · Electron + Python

---

💼 Production work (fintech / NBFC & manufacturing)

Direct credit bureau integration (CIBIL) — replaced a third-party aggregator with a direct integration: orchestration layer, PDF report generation, async processing at scale. ₹30L+/year saved.

Co-lending module (CLM2) — built the disbursement workflow that enabled co-lending operations: ₹14Cr+ actually disbursed, supporting a projected ₹70Cr+ pipeline.

EKS migration — moved 20+ microservices to Kubernetes on EKS, cutting infra costs by ~40%.

Billing job optimization — took a critical billing process from 6 minutes → 11.2 seconds (~32x faster) by eliminating sequential awaits and redundant DB round-trips.

Multi-vendor data fetch service — a single API layer over multiple data vendors with rate limiting, background jobs, and webhook callbacks.

Enterprise workflow portals — two greenfield systems for a large auto-components manufacturer: a CRM complaint portal (~74 endpoints, DB-configured approval workflows, per-request RBAC) and a warranty claims portal built as a pure table-driven state machine with idempotency keys, gapless claim numbering, and 3-system JWT federation.

---

✍️ Elsewhere

I write real backend debugging stories on LinkedIn — production war stories with the actual fixes.

---

!GitHub stats
!Top languages

📫 [[email protected]](mailto:[email protected])

Collapse