Field notes · an ascent through production AI

Fast, cheap, invisible. That's my bar for every layer of AI I ship.

Not just voice. Courier models on 3M+ shipments, RAG that finds the one paragraph that matters, agent harnesses that catch their own mistakes, a clinical platform built end to end. I train it, serve it, and grind the cost and latency down until the AI just quietly works — Bigship in logistics, freelance in healthcare.

CurrentlyAI Engineer, Bigship Technologies

ContractOpen to select engagements

BaseIndia (IST)

Book a call Begin the ascent

sys·log--:--:-- istvoice-agent · 100+ concurrent · 100k+ calls/day · 3k+ enterprise accounts

BM·01Selected work

The foothills. Every system here carries production scars — and the numbers to prove it.

Fig. 01 — call audio, 24 h window

Bigship Technologies

2025 — present

employment

LiveKit
Gemini
Deepgram
ElevenLabs
Python
FastAPI

Voice infrastructure for enterprise logistics

Bigship's clients needed outbound and support calls handled at a volume no human team could cover. I engineered the production voice stack end to end — LiveKit transport, Gemini reasoning, Deepgram STT, ElevenLabs TTS — with the unglamorous parts that make it real: voicemail detection, DTMF handling, dynamic call routing.

Real-time agent pipeline across STT, LLM, and TTS with dynamic routing per client
Voicemail detection and DTMF so the agent behaves correctly on real phone networks
Gemini-powered logistics chatbot with function calling across six tracking APIs, plus RAG over policy documents
Document AI pipeline (Gemini Flash) classifying invoices with fault-tolerant incremental saving

Measured

Concurrent calls100+

Daily call volume100k+

Enterprise accounts3k+

Latencycut 45%

Response accuracyup 80%(voicemail detection, DTMF, routing)

Invoice classification2,000+/day(~$0.005 per document)

Fig. 02 — roc — success classifier

Bigship Technologies

2025 — present

employment

LightGBM
Python
FastAPI
Pandas
scikit-learn

Courier recommendation engine on 3M+ shipments

Choosing the wrong courier costs money twice: late deliveries and returned-to-origin parcels. I built the ML system that ranks couriers per shipment — a dual LightGBM pipeline trained on 3M+ shipment records, served behind FastAPI with multi-objective ranking.

Delivery-time regressor and delivery-success classifier trained as a dual LightGBM pipeline
Multi-objective ranking service weighing speed, success probability, and cost per shipment
Feature engineering and ETL over 3M+ historical shipment records

Measured

Training data3M+ shipment records

Success classifier0.927 AUC-ROC

RTO recall81%(catching returns before they happen)

Delivery-time error±1.37 days MAE

Fig. 03 — a query walks the index

Zudu.ai · contract

Oct — Nov 2025

contract

LiveKit
Twilio
Plivo
PostgreSQL
JWT
Chargebee
Azure

Multi-tenant voice backend, 43 indexes deep

Zudu's AI calling platform needed a backend that could orchestrate dozens of simultaneous calls across tenants without queries melting down. I architected the LiveKit + Twilio/Plivo backend and then went through the database with a profiler, index by index.

Multi-tenant orchestration for 50+ concurrent AI calls over LiveKit with Twilio and Plivo trunks
43 targeted PostgreSQL indexes from systematic query profiling
11-endpoint analytics REST API with JWT auth, Chargebee billing, Azure Blob archival

Measured

PostgreSQL query timecut 90–95%(43 targeted indexes)

Webhook latency40 ms

Concurrent AI calls50+(across tenants in production)

Fig. 04 — cold start, log scale

Scicom.ai · contract

Oct — Dec 2025

contract

LiveKit Agents
Groq
LLaMA 3.3 70B
Deepgram
Pinecone

Real-time RAG voice avatar, 3.9 s → 378 ms

A voice avatar that takes 4 seconds to say its first word is a demo, not a product. I deployed Scicom's production RAG avatar on LiveKit Agents with Groq-hosted LLaMA and Pinecone retrieval, then attacked the cold start until it was imperceptible.

Production avatar pipeline: LiveKit Agents, Groq LLaMA-3.3-70B, Deepgram Aura TTS, Pinecone serverless retrieval
Cold-start elimination via connection pre-warming and lazy-init ordering
Thread-safe session handling for concurrent conversations

Measured

Cold-start latency3.9 s → 378 ms(10.4× reduction)

Sessionsconcurrent, thread-safe

Fig. 05 — one clinic week

Laser Vascular Center · contract

May — Jul 2026

contract

FastAPI
PostgreSQL
psycopg2
React 18
TypeScript
TanStack Query
Radix UI

Clinical scheduling platform, no ORM in sight

A multi-office clinic ran staff scheduling out of a Google Sheet maintained by 3–4 people. I replaced it with a full platform: FastAPI and raw-SQL PostgreSQL on the back, React 18 and TypeScript on the front, Google OAuth and three-tier RBAC holding the doors.

60 REST endpoints across 11 routers on FastAPI with raw SQL over psycopg2 — no ORM, every query visible
Multi-tenant, 16-table org-scoped schema with a 4-rule conflict engine and weighted candidate ranking for shift assignment
React 18 + TypeScript frontend on Vite, TanStack Query v5, Radix UI, across 13 role-gated pages
Idempotent Google Sheets time-off sync and an append-only audit log with before-and-after diffs

Measured

Scheduling headcount3–4 staff → 1 operator(~70% fewer people in the loop)

Designed capacity200+ staff(across multiple offices)

API surface60 endpoints(11 routers, no ORM)

Fig. 06 — the tool loop

Open source

2025

open source

Python
FastAPI
Next.js
Docker

github.com/UtkarshTiwari07/AI-Customer-Support

A support agent that can't be sweet-talked

Refund decisions are exactly where an LLM should not improvise. This containerized support agent runs a raw function-calling loop — no framework — against a deterministic policy engine, so prompt injection can change the conversation but never the outcome.

Raw LLM function-calling loop with a deterministic, injection-resistant policy engine
Next.js chat UI plus an admin dashboard that exposes the agent's reasoning trace
Single-command docker-compose deployment

Measured

Test suite80 pytest tests

Policy decisionsdeterministic(the LLM proposes, the engine disposes)

Also built · HumanizeResumeAI, multi-agent resume optimizer serving 30+ concurrent users · SnapDetect, multi-model image analysis — YOLOv5x, Mask R-CNN, OCR

BM·02Live demo

First ridge, where the machines start talking.

I build voice agents for a living. Talk to one.

This opens a live LiveKit session into the same agent stack I deploy for clients — speech-to-text, LLM, text-to-speech, all real-time. Allow the microphone and say hello.

Field recorder 01status: standby

Microphone permission required · audio only · nothing is stored

BM·03Expedition log

Camps along the route, in order of ascent.

001Feb 2025 — present

AI Engineer · Bigship Technologies

Promoted from AI-ML intern. Voice infrastructure, courier ML, document AI for a logistics platform.

full-time

002Oct — Nov 2025

Backend AI Voice Developer · Zudu.ai

Multi-tenant voice backend, PostgreSQL optimization, analytics API.

contract

003Oct — Dec 2025

AI Engineer · Scicom.ai

Real-time RAG voice avatar on LiveKit Agents.

contract

004May — Jul 2026

Full-Stack AI Engineer · Laser Vascular Center

Clinical staff-scheduling platform: FastAPI, raw-SQL PostgreSQL, React.

contract

Base camp · B.Tech, Computer Science & Engineering (AI-ML), Uttaranchal University, 2021 — 2025 · Oracle Cloud Infrastructure AI Foundations (2025) · Deep Learning with PyTorch — IBM (2024)

BM·04Toolbox

Pack weight. Only what has earned its place.

LLMs & agents

GPT · Claude · Gemini · LLaMA · RAG · function calling · agent loops · multi-agent orchestration · LLM-as-judge evals · MCP · LangChain · LangGraph · CrewAI · fine-tuning

Voice & real-time

LiveKit · Deepgram · ElevenLabs · Twilio · Plivo · WebSockets · VAD · DTMF

ML & data

PyTorch · LightGBM · XGBoost · scikit-learn · Pandas · NumPy · Hugging Face · ETL pipelines

Backend & MLOps

FastAPI · Flask · Docker · MLflow · Pydantic v2 · REST · microservices · CI/CD

Data stores

PostgreSQL · SQLite · Pinecone · semantic search · sentence transformers

Frontend & cloud

TypeScript · React · Vite · TanStack Query · Tailwind CSS · AWS · GCP · Azure

No proficiency bars, no percentages — a tool either shipped something real or it isn't listed.

BM·05About

The climber.

Fig. 07 — field portrait

Systems, summits, and stars.

I'm an AI engineer because production is where the interesting problems live: latency budgets, concurrency, the long tail of ways a system can fail. Voice was my proving ground — when a call drops mid-sentence, there's no hiding behind a demo.

Away from the terminal I'm drawn to mountains and the night sky. I studied computer science in Dehradun, at the edge of the Himalayan foothills, and kept the habit of looking up. This site is built from both loves: the daylight theme draws topographic contours from seeded noise, and night mode plots Orion from real J2000 coordinates. Deterministic, tested, no templates — the way I like my systems.

Reading terrain, respecting constraints, moving anyway. It's the same discipline whether the summit is a mountain or a launch.

BM·06Contact

Above the Kármán line the air is gone, but a clear signal carries forever.

The door is open.

Hiring for AI engineering, or have a system that needs to hold up in production? Book a call or write to me — I reply within 48 hours.

Emailutkarshtiwar89@gmail.com

Book a callcal.com · 30 min

ResumePDF, one page

GitHubUtkarshTiwari07

LinkedInutkarshtiwariai