Software engineer focused on AI infrastructure and developer tools. MS in Computer Software Engineering from Carnegie Mellon (Dec 2025), two years prior at Siemens Digital Industries Software. Mostly Python + FastAPI, deployed on AWS.
Portfolio → v0-simple-spa-deployment.vercel.app
- LLM serving infrastructure — caching, routing, rate limiting
- Retrieval systems — hybrid search, reranking, agentic tool use
- Developer tooling for AI workflows
Self-hostable mini autonomous SOC — one docker compose up, zero API keys. Ingests GitHub / AWS / Okta / Falco telemetry into a Neo4j security graph, then runs parallel Pydantic-AI agents (local Ollama) that triage the noise and reconstruct a supply-chain breach into a MITRE ATT&CK kill chain, with one-click response actions behind a human-approval gate and a SHA-256 hash-chained audit log. Demoed against real public attack data (Splunk Attack Data + a real lodash CVE): auto-closes ~84% of alerts and rebuilds an 8-stage incident end-to-end in ~2 min on CPU.
Python/FastAPI middleware between Retell voice-agent webhooks and Salesforce. HMAC-verified ingress, Postgres-RLS multi-tenant isolation, ON CONFLICT idempotency, Presidio PII redaction before any persist or CRM write, pgcrypto-encrypted OAuth refresh tokens, tenacity-retried CRM upserts, append-only audit log via REVOKE. Pluggable CRMAdapter Protocol (Salesforce reference impl + ServiceNow stub, mypy --strict). OpenTelemetry → Jaeger with seven named spans; inbound hydration under a 1.8 s budget for Retell's 2 s SLA.
Multi-instance LLM reverse proxy. Two-tier semantic caching (L1 LRU + L2 Redis embeddings, 15–73% hit rate), consistent-hash routing, circuit breakers, atomic rate limiting, request coalescing. Deployed on AWS via Terraform (ECS, ElastiCache, ALB) with OpenTelemetry tracing.
Full-stack RAG pipeline, no LangChain. Hybrid retrieval (pgvector + Postgres FTS + Cohere reranking via Reciprocal Rank Fusion), structural-aware chunking across 6 document formats, agentic tool calling, SSE streaming.
CLI coding agent built with Pydantic AI, Rich, and prompt-toolkit. Token-by-token streaming, runtime model switching across 3 providers, slash commands with tab completion, persistent history, 3-tier safety model.
Distributed systems learning lab — 40+ projects working through caches, message queues, write-ahead logging, circuit breakers, P99 latency tracking, and related primitives.
Languages: Python, TypeScript, SQL Backend: FastAPI, Pydantic, Pydantic AI AI/ML: OpenAI, Cohere, pgvector, sentence-transformers Infra: Docker, Redis, PostgreSQL, Terraform (AWS ECS / ElastiCache / ALB), Prometheus, Grafana, OpenTelemetry Graph / agents: Neo4j (Cypher), Ollama, MITRE ATT&CK, Sigma rules
- LinkedIn — kshitijkakade307
- Email — [email protected]
- Mountain View, CA