Piyush A.

Production AI agent systems: multi-agent orchestration, RAG grounding, and realtime voice agents. FastAPI, Python, GCP. Top Rated with 100% Job Success, 65 jobs and 835+ hours delivered on Upwork. I ship AI features that survive contact with real users — not demos. Most of my work is agent orchestration, retrieval that stays grounded, and sub-second voice pipelines, plus the guardrails, evals and observability that keep them reliable in production. WHAT I'VE SHIPPED RECENTLY Agentic analytics engine (in production, large tech client) A FastAPI service that answers analytics questions in plain English and cites evidence for every answer. Knowledge map of 57 metrics, 14 dimensions and 35 relationships, version-controlled as the source of truth. A constrained enum resolver means the model can only resolve metric IDs from a closed vocabulary, so hallucinated metrics are structurally impossible; vague queries trigger a clarifying question instead of a guess. Four agents — Planner, Researcher, an adversarial Critic and a Synthesizer — with schema-constrained tool calls. Looker and BigQuery underneath with grain-safe aggregation, freshness anchoring, trace IDs and an offline eval harness. cAlla — AI voice calling platform (live, UK outbound) Real inbound and outbound phone calls that qualify, book and follow up. Sub-second turn latency: Twilio Media Streams to Deepgram to Gemini 2.5 to Cartesia and back. Python/FastAPI realtime worker over WebSockets, streaming PCM and mu-law with resampling between telephony and model rates. Barge-in, tunable VAD endpointing, mid-call function calling, per-turn latency telemetry, pgvector RAG over uploaded documents. Multi-tenant with RBAC, audit logs, DNC handling and GDPR export. Singularity Learn — AI teaching agent (live) Curriculum-aware lessons with pedagogical guardrails, multi-provider routing across Gemini, Claude and GPT, 56+ interactive tools triggered mid-lesson and streamed to React over SSE, plus streaming voice and multilingual TTS. An earlier client's custom LLM + diffusion deployment cut their infrastructure bill by roughly $5,000/month. HOW I WORK You share the goal and the deadline. I come back with a day-by-day plan, success metrics and a fixed price. Reliability, speed and clean handoffs. STACK Python, FastAPI, Go, Celery, Redis, PostgreSQL, pgvector. Multi-agent orchestration, RAG, fine-tuning (SFT/DPO/GRPO), guardrails and prompt-injection resistance. WebSockets, SSE, STT/TTS/VAD. GCP (Cloud Run, Vertex AI, BigQuery, Looker), AWS, Docker, vLLM, TensorRT. Tell me what you're building and I'll tell you exactly how I'd approach it.