PracticeIndependent · one-person studio
BasedGhaziabad, India · Working worldwide
FocusVoice AI · On-device ML · LLM systems

AI that has to
work the
first time.

01 Live
Voice
Intelligence
Pipecat · LiveKit · Bolna
ElevenLabs · Deepgram
Sarvam · Plivo · FastAPI
02 Shipped
On-device
Models
Qwen2.5-0.5B fine-tuned
5-round training pipeline
Android · Kotlin · Offline
03 Next
RAG &
Retrieval
LangChain · Pinecone
Domain-tuned chunking
Custom reranker · Next.js
04 End-to-end
Full-stack
AI Products
FastAPI · Next.js · Android
Multilingual · HuggingFace
Hindi · English · Marathi

I build AI systems for companies that need them to actually survive contact with real users — not collapse the first time someone speaks in Hinglish, drops the call, or forgets the Wi-Fi.

Scroll
01 — Selected work

Things I've built and shipped.

001 / 05

Hospital Voice Agent

2025In production

Outbound AI calling system for a hospital network — calling discharged patients on real Indian SIM numbers, in their own language. Stack-compared Pipecat, LiveKit and Bolna against the same test scenario before committing. ElevenLabs for TTS, Deepgram for transcription, Sarvam for Hindi STT, Plivo for telephony. Unit economics close at ₹1.20–1.30 per minute at scale. That's the hard part nobody writes threads about.

PipecatLiveKitElevenLabs DeepgramSarvamPlivoFastAPI
Real call recordingHindi · Live SIM
002 / 05

Appointment Booking Bot

2025Live

A 24/7 conversational appointment agent built on LiveKit and Pipecat. Reads real scheduling data, checks availability across providers dynamically, handles rescheduling and cancellations — all in voice, all in real time. Not a decision tree with a voice skin. An actual reasoning agent that books appointments.

LiveKitPipecatElevenLabs DeepgramScheduling APIFastAPI
Demo call recordingBooking flow
003 / 05

MedExtract — Fine-tuned Model

2026HuggingFace

Qwen2.5-0.5B fine-tuned across five rounds on real Indian clinical transcripts — Hinglish, doctor shorthand, dose changes buried mid-sentence. Extracts eleven structured fields per transcript with precision that surprised the reviewing doctors. Small enough to run on a ₹25,000 tablet. The kind of fine-tuning where you're writing custom post-processors between every round because the failure modes are always new.

Qwen2.5-0.5BFine-tuning Medical NLPHuggingFacePython
medextract-v5 · inference · on-device
HuggingFace ↗
> "Patient chest pain, BP 140/90, metformin 500→1000mg BD, review 2 weeks"
{
  "symptoms":    "chest pain",
  "bp":        "140/90 mmHg",
  "medication""Metformin",
  "dose_change":"500mg→1000mg BD",
  "followup":  "14 days"
}
View on HuggingFace ↗
004 / 05

MedExtract — Android App

2026Delivered · In daily use

The model, wrapped in a full Android product. Runs entirely offline — no cloud, no subscription, no data leaving the hospital. Doctors dictate, the app extracts vitals, medications, dosage changes, diagnoses and follow-ups in real time. Also handles OCR on lab reports, vital history tracking, and report summarisation. Shipped as a branded APK, in daily clinical use.

Android / KotlinOn-device ML OCRVitals trackingOffline-first
MedExtract · Android · On-device Opens Drive
Next build · In progress

VenueBot — not your
average RAG.

Most RAG systems return the nearest neighbor. VenueBot will return the right answer. A discovery engine that reasons about constraints across turns, narrows results without restarting, and knows when to say no instead of hallucinating a venue that doesn't exist.

01 — Retrieval

Domain-tuned chunking and a custom reranker built for venue-specific jargon. Not off-the-shelf similarity.

02 — Conversation

Memory that narrows the result set across turns — "but outdoor" filters, it doesn't restart.

03 — Honesty

Confidence thresholds. When the system doesn't know, it says so. No hallucinated venues.

Interested in this project? Let's talk →
02 — What I build

Clear on the craft,
clear on the no.

What I build

  1. i.Voice agents that work on real phone lines — not just WebRTC demos with perfect Wi-Fi.
  2. ii.On-device models that run offline, respect privacy, and don't need a $400/mo API budget.
  3. iii.RAG systems tuned for your domain — retrieval that finds the right thing, not the nearest thing.
  4. iv.Full-stack AI products — FastAPI backends, Next.js frontends, Android apps, all of it.
  5. v.Multilingual systems for Indian markets — Hindi, English, Marathi, and what the next client needs.

What I won't build

  1. i."AI-powered" wrappers that quietly pipe everything to ChatGPT.
  2. ii.Investor-demo magic tricks that fall apart at user #3.
  3. iii.Projects where the spec is "make it more AI."
  4. iv.Month-long discovery phases before a single line of code.
  5. v.Hype. I'll leave that to LinkedIn.
03 — Approach

One builder. No theatre.

"I take on one or two serious projects at a time. I send working code, not decks. I don't disappear for a week when something breaks. If you want a thirty-person agency — we're not a fit. If you want a senior builder who'll actually ship your thing — we probably are."

— That's it. That's the pitch.
01

Scoped in days

You send a paragraph. I send back a real proposal with cost breakdowns, stack reasoning, and a timeline I'll actually hit. Usually in 48 hours.

02

Built in the open

You get the repo from day one. Weekly written updates, Loom walkthroughs when they help. No black boxes, no big reveal at the end.

03

Shipped, then supported

Handover isn't a zip file. I stay for the real-user edge cases and document things your next dev can actually read.

04 — Let's talk

Got something real
to build?

hello@piyushtyagi.work
Twitter / X Upwork GitHub LinkedIn Book a call