Sampreeth Avvari
Resume

Bay 018 projects · 6 articles

AI Engineering.

LLM systems built for a dental practice.

I build LLM pipelines for a dental practice that fill a schema, quote their evidence, keep patient fields out of logs, and leave the final call to a person.

Scroll to walk the bay

How I work

  • Built for a dental practice. Doctors, coordinators and patients read what these systems write, so there is no room for confident nonsense.
  • The model fills a schema. It fills a validated JSON schema and quotes the transcript, with a timestamp, for every rationale.
  • The code does the math. Gemini returns integers and gate answers, and plain Python turns them into the score.
  • Patient fields stay out of logs. The forms logger drops answers, names, emails and phones by name and by shape.
  • A person decides the match. The record matcher refuses when unsure, and a coordinator's link decision always wins.
  • The prompt is the small part. The real work is schemas, retries, fallback models, audit trails and where a human stays in the loop.

What I bring

Structured outputs, evidence rules, and humans where it matters.

  • Structured output, enforcedSchemas with validation, retries and fallback models. The model fills a contract.
  • Evidence behind every scoreEvery score quotes the transcript with a timestamp, or it does not get said.
  • Patient data kept out of logsLoggers that drop patient fields by name and by shape, and one service holding the practice system's credentials.
  • Humans where it mattersUncertain matches wait for a person. AI proposes; people approve anything irreversible.
Paused

Enterprise Search (RAG)

An internal search over Hybridge documents that cites the exact passage, filtered by who may see what. Paused since October 2026.

  • Two searches in one Postgres, pgvector by meaning and full text by exact words, reranked to 8 passages
  • Conflict rules: superseded versions drop out, then authority, then effective date
  • Group access is a SQL condition inside both searches, so a restricted passage never becomes a candidate
Guided tour
In production

Doc Coach: Consultation QA

Every Zoom consult scored by Gemini on seven clinical criteria; in October the median report landed 2.7 minutes after the transcript.

  • 471 consults scored live from May 19 to Oct 28 across three offices
  • Three layer doctor identity: Zoom participants, then transcript speakers, then names Gemini pulls out
  • Truncation and schema failures handled apart; no second scoring attempt needed in October
Guided tour
In production

NPC Coach: Call Coaching

Finds the new patient calls among thousands of recordings and grades each against the playbook, with a quote behind every criterion.

  • 690 real calls graded by October 29; in October the median grade came 14 minutes after the call ended
  • Hard gates show as a flag beside the score: a call can score 84 and still carry a missed gate
  • The model returns integers and gate answers; plain Python does the 40/40/20 math
Guided tour
In production

CDF: Gemini reads the patient first

Gemini writes the doctor's page from the intake survey and reads the panorex under the founder's own rules; a doctor makes the call.

  • A ruleset written from four cases the founder walked through, each rule citing its case and timestamp
  • The model lists the teeth, and plain code recomputes the score
  • Every Gemini call has a way out: retries on the same model, then newer models, inside a time budget
Guided tour
In production

Marketing Data Warehouse

A BigQuery warehouse joining ten marketing sources, ad spend to treatment value, documented so LLMs read a guide first.

  • All 657 staging and marts columns described, plus a data guide written for LLMs to read first
  • 98.6% of lead to patient matches come from the link staff already keep in Monday, as of October 29
  • A weekly Gemini eval on 22 golden questions scored 21 of 22 on October 26
In production

Marketing Data Connector

A Claude connector over the warehouse: sign in with Google once, and query only the data an owner ticked.

  • Live three hours after the first connector code on October 23
  • A guard points every table at the asker's own dataset of views before anything runs
  • 480 tool calls from 3 people in its first seven days
Design complete

CRM intelligence layer

Duplicate patients proposed by an LLM with mandatory human review, plus structured extraction of free text activity logs.

  • No merges on its own. Married couples share phone numbers
  • Every parse carries a confidence score and source reference for reruns
  • Consent tokens extracted into an append only ledger
Live

Portfolio chatbot

The assistant in the corner of this site, on Cloudflare Workers AI and running on the free tier.

  • Free tier serving on Workers AI, with no API keys and no vector database
  • Context as code: site copy compiled into the worker at deploy time

The writingAll posts

The CEO's consult checklist now grades every Zoom consultMay 2026 · 5 min read Every new patient call graded, with the quote behind each scoreJun 2026 · 5 min read Gemini reads the patient first, and the clinic waits for both reportsJul 2026 · 5 min read Ten marketing sources pulled at 6 AM and mailed at 6:30Oct 2026 · 5 min read A company search that cites its sources or says it does not knowJun 2026 · 5 min read JobPilot: the open source app that ran my job huntJun 2026 · 5 min read

End of the bay

Where next.

Get in touch Every project Resume The station