I architect agentic
systems & the infra
that survives production.

A healthcare assistant has to do two hard things at once: answer accurately from a large, changing medical catalog, and hold a natural realtime voice conversation — without overlapping speech, without leaking system internals into the transcript, and without ever inventing a patient's history or booking a lead that shouldn't exist. Most stacks solve one surface and rebuild everything for the other.

One Mastra agent runtime backs both surfaces. Retrieval runs over a namespaced Pinecone knowledge base (treatments, hospitals, doctors, listings, and user-uploaded medical PDFs) with an online ingestion pipeline. Persistent memory lives in PostgreSQL + pgvector — structured working memory, observational memory, and semantic recall — hydrated per thread from the CRM/EHR (leads, patients, calls) via LLM summarization, so the agent knows a user's lead stage, doctor, and history with no extra lookup. Ten agentic tools cover symptom checking, semantic search, document Q&A, and server-validated appointment booking. Voice bridges to Gemini Live over WebSocket behind a turn state machine with server-side VAD barge-in; every turn on both channels is traced in Langfuse with sampled LLM-as-judge medical scorers.

• Text and voice share one runtime, thread, and memory — a phone/CRM caller and an app user with the same number resume the same conversation
• Booking is validated server-side before any lead exists — real mobile numbers only, doctor and hospital IDs only from search results — so it never fabricates a lead
• Removed a 2–3s stall before the voice assistant's first words with Gemini's asynchronous function calling for memory writes
• Gated the LLM prompt-injection check to skip short benign turns (−1–1.7s each); agent prompts, models, and tools are DB-driven and hot-swappable without a redeploy

One codebase serves both a streaming chatbot and a realtime voice assistant with shared memory and patient context, medical-safety evals on every sampled turn, and runtime-tunable agents — with a 2–3s voice greeting stall removed, 1–1.7s cut from short turns, a ~3s RAG cold start gone, and tracing fixed for up to a quarter of traces.

AI writing tools generate fluent but off-voice content, and lean on models that don't reliably return valid structured output — so the backend either bloats with brittle glue or breaks the moment you swap the LLM. The harder problems are staying on-voice, keeping JSON trustworthy, and staying portable across providers.

A 6-agent Mastra pipeline (Persona Analyst → Onboarding → Trend Research → Content Generator, plus a live Persona Chat and an AI co-writer) turns a user's own posts into 5–10 trend-anchored, personalized briefs. A deterministic Writing-DNA engine extracts 16 voice metrics at zero LLM cost. The whole app is provider-agnostic — every agent resolves its model through one factory, switchable between Google Gemini and Ollama Cloud by an env var — and structured output is enforced in four cheap-first layers (native JSON mode → prompt rule → local repair/normalize → at most one repair call) so no call depends on schema-constrained decoding. Trend intelligence classifies each creator into one of 14 industry domains and fans out across Tavily, Hacker News, 66 curated RSS feeds, and Google News, with cross-source dedup, a two-tier cache, and an LLM-free fast path when relevance scoring is confident. Feedback — explicit ratings plus implicit behavior — feeds a recency-decayed learning loop, weighting reported post performance highest. Turborepo monorepo: 3 packages, 11 REST modules with Swagger, 13 MongoDB collections; a Next.js 14 app with real-time chat, an AI co-writing editor, and a 7-signal AI detector plus humanizer.

• Provider-agnostic LLM layer — swap Gemini ↔ Ollama Cloud with one env var, no code change
• Never hand the SDK a response schema — enforce JSON in four cheap-first layers so unreliable models still return valid output
• Deterministic Writing-DNA fingerprint — 16 voice metrics at zero LLM cost, 100% reproducible
• Real trend sources over LLM-generated trends — 66 RSS feeds across 14 domains, with an LLM-free fast path when scoring is confident

A live content-intelligence platform that generates on-voice briefs from real-time trends, learns from every rating and reported result, and runs on either a hosted or self-hosted LLM.

Legal contracting on the web is mostly PDFs and DocuSign. Litt Law needed real-time concurrent editing, AI-drafted clauses backed by retrieval, hard uptime guarantees, and reliable payments — all at the throughput a production legal platform actually faces.

Node.js + Express microservices over gRPC. PostgreSQL with Drizzle ORM as the system of record, Redis for caching. 40+ JWT-authed REST APIs. RAG pipelines use custom embeddings on Pinecone. Real-time collaboration (presence, concurrent editing) is delivered over SSE. The Next.js (SSR) frontend uses React, Redux, Tailwind, and Tiptap. Razorpay and Cashfree handle payments. A usage-based credit system meters AI chat on top of the payment flow, and Prometheus metrics with Grafana dashboards monitor the services.

• Custom embeddings over off-the-shelf — better domain fit on Indian legal corpus
• SSE for collaboration instead of WebSocket — simpler ops at the scale we needed
• 95%+ test coverage on the backend; nothing merges without a green eval path
• Two payment providers with idempotency keys — Razorpay + Cashfree at 100% reliability

99.9% uptime, 2.3× throughput improvement, 48% latency reduction. Scales to 1,500+ concurrent users with 500+ monthly signings across two payment providers.