the whole wall

What I've built, and what it did

What each product does for the people using it, what I built, and a simple picture of how it works. Kindred AI products are built with a team of product, design and engineering colleagues, and I say which parts I owned. The side projects are things I build outside work.

Kindred AI

Senior Product Engineer since September 2023, working remotely. Kindred Workplace and Kindred Practice are the products I've spent most of my time on; Kindred Core and Agent Ops are the shared AI and tooling underneath them, and I've also delivered Medimap and the Stellantis Skills Platform.

Kindred AI · emotional intelligence at work Measure how a team works together. Then help everyone get better at it.

Kindred Workplace

Current

Senior Product Engineer · AI coach backend, Analyze, Voice ID, Slack and Teams

The costliest problems at work live between people: the six meetings that replace one honest conversation. Kindred Workplace measures emotional intelligence across a company and coaches every person to grow it, in the app, by voice, and inside Slack and Teams.

  • 8,300+ people on the platform
  • 21 companies using it
  • 90 min conversations Analyze turns into coaching feedback

What I built

  • A coach the team can trust. The multi-agent backend behind Paige, the AI coach: a supervisor routes each message to specialist agents, replies stream back, and an eval suite of golden conversations alerts the team when quality drops.
  • Analyze. Turns a recorded conversation of up to 90 minutes into coaching feedback.
  • Voice ID. Recognises the member's own voice, with per-user usage limits that keep AI cost predictable, and consent and GDPR deletion built in.
  • Coaching where people already work. Paige inside Slack and Microsoft Teams with scheduled check-ins, and the ElevenLabs voice agent connected securely to live product data.
kindredworkplace.com ↗

TypeScript · NestJS · Postgres · Claude · ElevenLabs voice · Slack & Teams

Kindred Workplace: measure, then grow

People are measured where they are, coached by an AI that remembers them, and companies see the change.

Where people are
The app web and phone
Slack + Teams coaching in the flow of work
A real conversation recorded, with consent
Measure
EQ assessment an emotional intelligence profile
Analyze recordings up to 90 minutes
360 feedback from peers, in rounds
Grow Paige remembers you
Paige, the coach chat and voice
Nudges + to-dos small steps each week
For the company
Team reports where the team stands
Adoption dashboards who's engaged, who's stuck
Kindred AI · formerly Mynd Therapists stay present with clients. The software handles the rest.

Kindred Practice

Live

Product and engineering owner · sessions, notes, scheduling, billing, AI

Therapists lose evenings to notes, scheduling and billing spread across a pile of tools. Kindred Practice puts sessions, clinical notes, calendar and payments in one place, with AI features that take on the paperwork. I own its product and engineering, working with product, design and clinical stakeholders.

  • 1,000+ therapists run their practice on it
  • 61% lower monthly AWS production bill
  • ~40% faster AI answers at p95, with ~30% lower LLM spend

What I built

  • Therapy without leaving the app. Live video sessions on LiveKit and AWS Chime, with Deepgram transcription.
  • The business side, handled. Scheduling, billing and clinical notes in one place.
  • AI answers therapists can rely on. Took retrieval (RAG) from prototype to production with eval-driven iteration, cutting p95 latency ~40% and LLM spend ~30% through prompt caching and model routing.
  • A leaner bill. Cut the monthly AWS production bill 61% by right-sizing instances and removing idle, duplicated and over-provisioned resources.
kindredpractice.com ↗

Django · Node.js · Postgres · LiveKit video · AWS Chime · Deepgram · AWS

Kindred Practice: from booking to a signed note

A client books, the session happens on video, and by the time the therapist closes the tab a draft note is waiting.

People
Client books, joins, pays
Therapist runs the practice
The session
Booking + reminders scheduling and reminders
Video room live video on LiveKit and AWS Chime
Transcript speech to text with Deepgram
AI does the paperwork Paige
Draft note in the therapist's own template
Ask Paige “what did we cover last month?”
The business
Billing invoices and payments
Review + sign the therapist has the final say

what it gives every product

  • Memory that knows when a fact was true
  • Speaker recognition across microphones
  • Usage metering and billing credits
  • Evals that can block a bad release
Kindred AI · shared AI platform Build an AI feature once, and every Kindred product gets it.

Kindred Core

In production

Memory, speaker recognition, metering and evals

Every Kindred product and partner needs the same hard AI pieces: memory, voice, billing and safety checks. Kindred Core is the multi-tenant platform that does them once, properly, so each product can ship AI without starting from zero.

  • ~2× accuracy on time-sensitive memory questions
  • 24/24 correct speaker decisions in testing
  • 24 simulated personas in every pre-release review

What I built

  • Long-term memory. AI agents remember what users shared and when it was true, with full erasure on request. Roughly doubled accuracy on time-sensitive questions.
  • The right speaker, every time. Tuned speaker recognition on real production audio so it identifies the right person across different microphones: 24 of 24 correct decisions in testing.
  • Ready for partners. Usage metering, billing credits and an audit log of every AI decision, with per-partner data rules and complete deletion of a person's records on request.
  • Tested before every release. Automated checks that can block a bad release, plus model-graded quality reviews across 24 simulated user personas.

NestJS · Postgres · pgvector · Claude · Langfuse · Inngest

Kindred Core: the AI every product shares

Products don't each build their own memory, voice and billing. They ask Core, and Core answers the same careful way for all of them.

Products
Kindred products Workplace, Practice and more
Partner companies each with its own data rules
Kindred Core
Memory what was shared, and when it was true
Speaker recognition the right person, any microphone
Metering + credits usage and billing per partner
Erasure a person's records deleted on request
Always on
Audit log every AI decision recorded
Evals can block a bad release

what it does for the team

  • Turns errors into fixes
  • Drafts support replies
  • Keeps help articles current
  • Blocks patient data and secrets
Kindred AI · engineering automation Bug reports in. Reviewed fixes out.

Agent Ops

In production

Coding agents, support agent and team tooling

Engineers were losing time to triaging errors and small tickets. Agent Ops hands that work to AI agents that investigate and write a plan; a person approves the plan, the agent opens the fix as a pull request, and a review agent checks it.

  • 30+ ready-to-review pull requests from agents each month
  • Every plan approved by a person before code is written
  • Days for new hires to ramp up, down from weeks

What I built

  • Error to pull request. Coding agents turn Sentry errors and tickets into ready-to-review pull requests, with a human approving each plan and a review agent checking every PR.
  • Support that knows when to ask. A support agent drafts customer replies and hands off to the team when unsure, and a pipeline keeps help articles up to date.
  • Safe by default. A scanner blocks patient data and secrets from agent changes.
  • Shared ways of working. The Claude Code skills and Cursor rules the engineering team builds with, cutting new-hire ramp from weeks to days.

Claude · Claude Code · Cursor · Sentry · Inngest · Langfuse

Agent Ops: from an error to a reviewed fix

Agents do the digging and the first draft; people make the call.

Something breaks
A Sentry error
A ticket
Agents
Investigate read the code, write a plan
Safety scan blocks patient data and secrets
Loop
A person approves the plan, before any code
Pull request checked by a review agent

Stellantis Skills Platform

Multi-rater skill assessments for 25,000+ employees in 14 languages. Delivered SAML single sign-on and manager-based permissions, and set up AWS production with idle servers sleeping to cut cost.

Medimap

Walk-in clinic finder covering 4,200+ clinics across six provinces. Rebuilt the platform after acquisition on Next.js and Payload CMS and launched online booking; revenue and users each grew 15% in two months.

Side projects

Independent products I design, build and launch outside my day job, under the Soonbuilt name.

Side project · web, iOS, Android, Mac, Windows One line in, a finished video out, scheduled to 15 networks.

TallyReel

Live

Side project · design, build, launch

Posting short video every day means a script, a voice, footage, captions and a schedule across a dozen apps. TallyReel turns one line into a finished short video and schedules it to 15 social networks, from the web, a phone, a desktop app, or straight from Claude and ChatGPT.

  • 15 social networks it schedules to
  • 52 tools in its remote MCP server for Claude and ChatGPT
  • 3 autonomy levels for the in-app AI agent

What I built

  • One line to a finished video. Script, voice, footage and captions from a single line, then scheduled to TikTok, YouTube, Instagram and 12 more networks.
  • Wherever you are. Web (Next.js), iOS and Android (Expo), and a Mac and Windows desktop app.
  • Works inside Claude and ChatGPT. A remote MCP server with 52 tools and OAuth, so people can create, schedule and review videos from their AI assistant.
  • An agent you set the leash on. An in-app AI agent with three autonomy levels, and per-video GPU cost tracking on Modal.
tallyreel.com ↗

Next.js · Expo · MCP · OAuth · Modal

TallyReel: one line in, posted everywhere

Start from the app, the desktop, or your AI assistant. TallyReel makes the video and posts it on schedule.

You
One line “How deep the ocean really is”
Where you ask
Web, phone, desktop Next.js, Expo, Mac and Windows
Claude or ChatGPT remote MCP server, 52 tools, OAuth
In-app agent three autonomy levels
Make
The video script, voice, footage, captions
GPU cost per video tracked on Modal
Out
15 networks TikTok, YouTube, Instagram…
Side project · web Practice out loud before the round that counts.

Round Zero

Live

Side project · design, build, launch

Most interview prep is a question bank; almost nothing lets you rehearse the real conversation. Round Zero is a live AI voice interviewer that adapts its follow-up questions as you answer, scores you against a rubric for the role and level, and runs coding rounds too.

  • Live voice interviews that adapt follow-ups in real time
  • By level role-specific rubrics calibrated by seniority
  • Coding rounds as well as conversation

What I built

  • A real interviewer, by voice. Built on LiveKit, it listens and adapts its follow-up questions in real time.
  • Fair scoring. Answers are scored against role-specific rubrics calibrated by seniority.
  • Coding rounds. Practice the technical part of the loop, not just the talking.
joinroundzero.com ↗

LiveKit · Voice AI · LLM evals

Round Zero: a live interviewer that adapts

Pick the role and level; the interview happens by voice; you're scored against a rubric made for that role.

Set up
Role + seniority
Rubric role-specific, calibrated by level
The interview
You speak live voice on LiveKit
The interviewer adapts follow-ups in real time
Coding round the technical part, too
After
Scorecard scored against the rubric
You read and edit the script before anything renders.
Post: pin a video to the week and it goes out on its own.

Smaller side projects

Two more independent products, both live.

Live

Blurt

Hold a key and talk, and clean text appears wherever you type, with on-device speech recognition in Rust for Mac and Windows.

blurt.soonbuilt.com ↗
Live

Maina

Voice cloning from 30 seconds of speech in English, Hindi and Urdu, with a built-in AI watermark and evals for word error rate and speaker similarity.

maina.soonbuilt.com ↗

Earlier

Work from Extendly, Zaytrics and university.