Dishank Jhaveri

AI Engineer — Voice Agents, Tool-Calling LLMs, and RL-Tuned Decision Systems
📧 dishankjhaveri@gmail.com · 💼 LinkedIn · 🐙 GitHub @dishank19

---

What I Build

💬 Voice AI agents that actually do work: answer calls, extract facts, take action
🧭 Agentic systems with reliable tool access (LangGraph + MCP) and guardrails
☁️ Production infra on GCP with clean observability and sane autoscaling

Ship real agents for real teams—safely, fast, and at scale.

---

Featured Work

PatientIntakeVoiceAgent
Production-style voice intake with turn-taking, info capture, slot-filling, and warm handoff.

Stack: Python · FastAPI · Pipecat · OpenAI · Cartesia TTS · Twilio · Docker
Repo: dishank19/PatientIntakeVoiceAgent

Google_Ads_Agent
Agent that researches keywords, adjusts bidding strategies, and manages ad groups autonomously.

Stack: Python · FastAPI · GAQL · LangGraph

leadgen-agent
Dental lead qualification with clinic lookup, appointment simulation, and automated notifications.

Stack: Python · Retell/Voice · Twilio · Livekit

Mistral7B_Finetune
Fine-tuned Mistral-7B on SHP dataset for AskHistorians-style long-form responses.

Stack: PyTorch · QLoRA · PEFT · HuggingFace Transformers

---

Currently Building

RL-tuned agents — Train policies in simulated funnels (dialog, keyword bidding, handoff) to optimize business KPIs instead of proxy metrics
Custom environments — Gymnasium/PettingZoo-style envs + reward models that mix goal attainment, latency budgets, and safety constraints
Safety & eval — Risk critics + rule gates; scenario suites; offline → canary → prod promotion

flowchart LR
  Logs[Prod logs] --> Offline[Offline dataset] -->|pretrain| Policy[Policy]
  Sim[RL Envs] -->|rollouts| Policy -->|actions| Sim --> Rewards[Reward + KPI] --> Policy
  Policy --> Gate[Safety Gates] --> Canary --> Prod

Loading

---

Skills

Core:          Python, TypeScript/JavaScript, SQL, Go, Java, C/C++
Agents:        LangGraph, MCP, tool-calling, RAG
Voice:         LiveKit, Pipecat, Twilio (SIP/PSTN), WebRTC, Cartesia TTS
RL:            Gymnasium, Verifiers, Stable-Baselines3, RLlib
ML/LLM:        OpenAI API, HuggingFace, PyTorch, Transformers, spaCy
Backend:       FastAPI, Node/Express, React, REST APIs
Cloud/Infra:   GCP (Cloud Run, GKE, Logging), Docker, GitHub Actions, AWS
Data:          PostgreSQL, MongoDB, S3/GCS, Pandas, NumPy

---

Experience

AI Engineer — Neurality AI · San Francisco
Shipped production voice agents (scheduling, clinical docs, insurance verification) on GCP; designed LangGraph workflows and MCP tool surfaces; built internal copilots that reduced manual ops.

ML Research — Graphite.io · Long-form style transfer (SELF-DISCOVER), large-scale evaluation, classifiers for formality/politeness/humor/simplicity

ML Engineer — Street Style Store · ETL, realtime analytics, recommender systems/segmentation, ROI-driven data tooling

Education: M.S. Computer Science, UMass Amherst · B.Tech Computer Science, VIT

---

Design Philosophy

# Voice agent loop with tool access and safety
while call.active():
    text = asr.listen_stream()
    state = dialog.update(text)
    action = planner.decide(state)              # LangGraph node
    if safety.blocks(action): action = planner.fallback(state)
    result = tools.execute(action, via="MCP")   # standardized, authenticated surface
    tts.say(render(result))

Building production agents with RL-driven performance? Let's chat: https://calendar.app.google/UMEFJTfYpFtCUXBG7