I build backend systems and AI infrastructure with a focus on correctness, reliability, performance, and data isolation.

Focus Areas: TypeScript · Python · PostgreSQL · Distributed Systems · Real-Time Systems · LLM Applications
I also contribute to open-source infrastructure, particularly around agent runtimes, execution reliability, authentication, and performance.

📍 India · Open to Backend, Platform & AI Engineering roles
📫 LinkedIn · amritraj4work@gmail.com

My strongest external contributions focus on agent harnesses and distributed actor runtimes:

Contributed fixes around agent execution reliability, validation, tooling, and cross-platform behavior:

• ✅ Merged Upstream (#7356): Added nearest-tool suggestions for unknown tool calls using Python's difflib matching to improve agent execution resilience.
• LLM Self-Correction & Retry (#7392, #7382): Captured structured Pydantic schema validation failures in the orchestrator and fed error reports back into the LLM context for real-time autonomous correction.
• Fail-Fast Orchestrator Validation (#7395): Added load-time validation for orchestrator output_keys to prevent silent graph execution failures.
• Cross-Platform Runtime Reliability (#7381, #7377): Resolved Windows path resolution issues for MCP servers and added retry logic for atomic file operations.

Contributions around runtime performance, authentication, and WebSocket protocol handling:

• $O(N) \to O(1)$ Request Lookup (#5635, Issue #5581): Profiled runner tunnel under high concurrency, isolated CPU bottleneck caused by linear array scanning in requestToActor, and refactored to an $O(1)$ Map lookup.
• External OIDC/JWT Infrastructure (#5572): Added external OIDC and JWT token verification inside onAuth lifecycle hooks with in-memory JWKS public-key caching.
• Inspector Protocol Fix (#5596): Fixed missing WebSocket protocol headers required by actor inspector client connections.

TypeScript · Node.js · React · WebSockets · PostgreSQL · Docker

A deterministic exchange simulator exploring orderbook data structures, fixed-point arithmetic, durable event journaling, crash recovery, and transactional settlement.

• Measured SkipList Speedup: A controlled 50K-resting-order benchmark reduced cancellation time from 37.5s to 29.8ms by replacing linear order scanning with a SkipList price index and an $O(1)$ order-ID lookup map.
• Domain Invariants: Implemented strict fixed-point integer arithmetic (PriceTicks & QuantityLots) to eliminate IEEE-754 floating-point drift, verified across 100K-operation differential tests.
• Crash Recovery & Reconciliation: Implemented durable event journaling with snapshot/replay recovery that survives abrupt SIGKILL termination, settling idempotently into an ACID double-entry PostgreSQL ledger with automated 3-way balance reconciliation.

Next.js 15 · PostgreSQL / pgvector (Supabase) · Gemini 2.5 Flash · SSE Streaming

A multi-tenant RAG application with database-enforced workspace isolation, validated tool calling, source citations, streaming responses, and ingestion idempotency.

• Database-Enforced Multi-Tenancy: Workspace isolation is enforced at the database layer inside a custom PostgreSQL match_chunks RPC (WHERE dc.workspace_id = target_workspace_id) across a shared HNSW index.
• Grounded Responses & Refusal: Retrieval thresholds and system-level rules trigger explicit refusals when relevant workspace context is unavailable, coupled with inline source citations.
• Validated Tool Execution & Idempotency: Schema-validated tool execution for side-effects; enforced SHA-256 content_hash constraints to guarantee idempotent document ingestion.

React 18 · Vite · TypeScript · Tailwind CSS · Zustand · Zod · Google Gemini

A technical prototype exploring audio transcription UX, structured clinical entity extraction, schema validation, and LLM-assisted workflow orchestration. (Not intended for clinical use.)

• Demonstrates browser-based audio capture, structured extraction into clinical schemas (diagnoses, medications, follow-ups), and self-healing validation loops routing Zod errors back to the model for real-time schema correction.

Python · Docker · Cypher · AQL · Stanford SNAP Dataset

A reproducible benchmark evaluating graph database engines (CognoDB Cloud, Neo4j AuraDB, Memgraph Cloud, ArangoDB, FalkorDB) under controlled workloads, standardized datasets, and explicit fairness/measurement caveats.

• Evaluates data ingestion, 1/2/3-hop traversals, aggregations, and concurrent read/write workloads with p50/p95 latency metrics.
• Explicitly documents architectural trade-offs, network-vs-local latency deltas, and resource constraint limitations across tiers.

• Yield Visualizer: Full-stack real-time US Treasury yield curve and 10Y-2Y spread inversion tracker built with Python (FastAPI), WebSockets, and canvas charts.
• FinUI Design System: High-density React component library monorepo (@amrit_16/core, Storybook 8, Tailwind tokens, tsup) built for financial interfaces.
• Financial Data & RBAC API | Swagger Docs: Backend REST API demonstrating role-based access control (Admin/Analyst/Viewer), automated JWT validation, and financial record processing built with NestJS and Prisma.
• Pen-Pulse | Live Demo: Earlier backend project exploring Cloudflare Workers, Prisma Accelerate connection pooling, and shared Zod validation.

Distributed Systems · Backend Architecture · AI Infrastructure · Real-Time Systems · Reliability & Fault Tolerance · Performance Engineering