Siddhant Tiwary - Software Engineer & AI Systems Architect

Experience

Roles & impact

Stealth AI startup

Remote · Jan 2026 – Mar 2026

Architected an LLM orchestration layer combining real-time biometrics (HealthKit/EHR) with knowledge graphs and semantic search; used pre-aggregated profiles to reduce compute while surfacing medical insights via tool-calling.
Integrated VAPI as a voice layer on the existing conversation engine with VAD-based turn detection without duplicating conversation state.
Built a stratified prompt-context layer and instrumented LangSmith tracing for tighter evals.
Led HIPAA-oriented security design for PHI and coordinated App Store and regulatory workstreams.

Bangalore · Mar 2025 – Feb 2026

Introduced Argo Workflows to orchestrate and track executions with configurable resources.
Deployed an in-cluster Harbor proxy cache to avoid upstream rate limits, speed up image pulls, and cut network cost.
Built a centralized migration service for schema and data migrations across multiple databases.

Bangalore · Jun 2022 – Mar 2025

Migrated production workloads from AWS Lambda to Kubernetes on spot instances, cutting latency ~68% and cost ~50%; used Ray for distributed workloads.
Designed and shipped a multi-tenant FIU module for bank account data aggregation, serving 5M+ users across 100+ clients.
Improved DynamoDB access with a Redis cache layer, reducing RCU use and eliminating read throttles.
Replaced an O(n) completion tracker with an atomic O(1) counter for distributed extraction jobs.
Stood up Kafka (streaming), Vector (batching), and ClickHouse for analytics and monitoring.

Selected work

Aug 2026 – present

46 LLM personas advance on an hourly world tick; each reads a heat-ranked feed and decides in character whether to post, reply, quote, like, or scroll past.
Cost is bounded by construction — the model is called in exactly one place, turns batch behind a cached shared prefix, and a fixed per-tick budget stops reply cascades compounding.
Python worker on Actions cron, Postgres, Next.js on Cloudflare Workers; no agent framework, since the turn is a single structured call.