Each point is a passage about my work. Color is the body of work.
  • Voice & speech-to-text21
  • Agents18
  • RAG / True Bible Answers9
  • Media apps5
  • Early work4

Member of Technical Staff at Speechify · Berlin, open to remote

Tim Dolenko

I build software that listens, and software that builds software.

Ten years of shipping products. Today I own the AI voice stack at Speechify end to end: a real-time speech-to-text iOS keyboard, its dictation SDK and the Rust backend. On the side I build autonomous agent harnesses that take a ticket to a tested, video-documented pull request.

years shipping products
10
commits on GitHub
16,000+
merged PRs at Speechify in 10 months
178
tickets delivered
511
tickets shipped by my agent pipeline
90+

Voice & speech-to-text

Speechify Voice Keyboard

An iOS keyboard you talk to. Streaming speech-to-text over WebSockets, a native QWERTY layout, Speak-to-Edit, emoji and AI autocorrect, all inside the tight memory budget of a keyboard extension.

users in the first 4 days
18.5K
words dictated in 10 days
100K+
cold launch
11 s1.7 s
empty-result sessions
30%<3%
dictation start
800 msinstant
offline survival of a session
7.5 s1 h
Voice Keyboard, light mode, in Reminders
Voice Keyboard listening to the iPhone microphone
Voice Keyboard, dark mode
  • Cross-process design: the keyboard and the app share one mic through an arbiter and talk over IPC throttled to 30 Hz.
  • Sessions survive bad networks: per-segment reconnect, a 2 / 6 / 15 s retry ladder, and no lost words when the connection drops.
  • Battery use measured with Power Profiler and brought to parity with Wispr Flow.
  • Five layouts (QWERTY, QWERTZ, AZERTY, Dvorak, Colemak), native haptics, 1,200+ emoji, iPad support.

Swift · SwiftUI · Keyboard extensions · AudioUnit · WebSockets · Rust (Axum, Tokio)

Voice & speech-to-text

Hybrid autocorrect

On-device SymSpell with bigram context for speed, plus an LLM pass for whole sentences. Benchmarked against real typo corpora before every release.

accuracy on 4,299 Wikipedia typos
71.7%84.8%
good words wrongly changed
1.41%0.55%
more typing after rollout
5.2×
LLM cost per day
$0.20
  • 200+ recorded probes of the native iOS keyboard, grouped into 12 correction types, to match its behaviour.
  • A 10K-word index keeps the keyboard extension under its memory limit.

Swift · SymSpell · Inception Mercury · Rust

Agents

speechify-pipeline

A Linear ticket goes in. A headless Claude Code agent works in its own git worktree, reproduces the bug on video, fixes it, drives the iOS Simulator through QA with recorded videos, reviews its own PR and ships a TestFlight build. I review the result from my phone.

tickets shipped
90+
parallel lanes
4
my merged PRs per month (10 → 33)
tickets shipped on the first production run
6 / 6
  • Ticket → worktree → bug reproduced on video → fix → Simulator QA and edge-case pass on video → AI code review → PR → TestFlight → Telegram report.
  • Leased simulator pool, per-lane locks, stuck-run detection and acceptance checks.
  • Controlled from Telegram: a triage agent answers questions, queues tickets and pauses or resumes lanes.
  • A second agent reads production logs, groups failed dictations by signature and files deduplicated tickets.

Claude Code (headless) · Python · Bash · xcrun simctl · ffmpeg · fastlane · Linear API · GitHub CLI · Telegram Bot API

RAG / True Bible Answers

True Bible Answers

An answer engine for Bible questions in 21 languages. Every answer quotes its sources word for word.

verses in 252 translations
6.9M
retrieval streams fused with RRF
5
verse commentaries generated
27K+
LLM-judge quality scores
62K
  • A Claude Haiku query planner predicts the phrases, verse references and authors an answer needs.
  • Five retrieval streams: chunk, HyDE and document vectors (Voyage, pgvector HNSW), full-text search and a verse-reference index, merged with Reciprocal Rank Fusion.
  • Claude Opus writes the answer with a corpus-search tool. Quality is tracked with Recall@k, MRR, nDCG@10 and LLM-judge scores across 30 versioned prompts.

Next.js 16 · React 19 · FastAPI · Supabase · pgvector · Voyage · Claude · Stripe

truebibleanswers.com

Agents

Octo

One agent that runs all my projects. Two hosts, a MacBook for anything that needs Xcode and an always-on VPS for everything else, each owning its own queue.

Claude sessions logged
8,368
projects under one runner
6
commits written by the agent on True Bible Answers
118
  • Heartbeat failover between Mac and VPS, and a deploy lease that allows one deploy at a time.
  • systemd workers, a SQLite activity database with full-text search, a dashboard and Telegram control.
  • A budget-aware scheduler turns the Claude usage left this week into a page count.

Python · SQLite · systemd · launchd · Tailscale · Telegram

Career

  1. 2025 – nowSpeechifyMember of Technical StaffAI Voice Keyboard, speech-to-text SDK, Rust backend, agent infrastructure.
  2. 2022 – 2026Axel SpringerSenior iOS Developer and Tech LeadThe new BILD app, Germany’s #1 news app. White-label SDKs for 3M+ monthly users. Subscriptions behind €6M+ a year.
  3. 2020 – 2022Spark NetworksSenior iOS DeveloperZoosk and 14 dating apps. 20+ modules, 92% test coverage on revenue-critical chat.
  4. 2018 – 2020WheeiOS EngineerA Silicon Valley family app from ex-Yahoo founders. Helped raise a $1M seed round.
  5. 2016 – 2018SispoFounderMobile agency. 15+ apps for clients like the NSW Department of Education and the UN World Food Programme.
  • Speaker at Media & Tech Berlin, 1,000+ attendees
  • 10+ articles and YouTube videos on iOS, CI/CD and async/await, 40K+ views
  • B.Sc. Computer Science. Product Management at Stanford.