Voice AI Beat Human Debt Collectors. Sinhala Is Wide Open

Voice AI agents are now doing a job that used to need a room full of people with headsets, and one YC company just put a price on it. CollectWise, a YC F24 startup, is hiring an AI Agent Engineer to build voice infrastructure for automated debt collection.

You probably can't take the job. It's onsite in New York and the listing says "US citizen/visa only." Read it as a market report instead.

---

📊 What the listing actually tells us

Job ads from small startups leak more than press releases do. Here's what this one puts on the record:

Fact	Value
Team size	8 people
Founded	2024 (YC F24)
Claimed run rate	$2M annualized, targeting $10M next year
Stated market size	$35B US debt collection
Their claim on quality	AI agents "already outperforming human collectors by 2X"
Salary band	$200,000–$300,000
Equity band	0.25%–1.00%
Core stack	Node.js, AWS, SQL, LiveKit

That 2X figure is the company's own claim in its own job ad, not an audited benchmark, so weigh it accordingly. But the shape of the business is hard to argue with: eight people, roughly $250K of run rate per head, running phone conversations that humans used to run.

Key takeaway: The interesting number isn't the salary. It's that eight people are servicing a call-heavy business at $2M ARR. Voice agents collapse the headcount curve of any operation whose product is "someone talks on the phone."

---

🎙️ The stack is smaller than you think

Look at what they're asking for. Not "PhD in speech recognition." They want 3+ years of backend or infrastructure engineering and 2+ years of voice AI, conversational AI, prompt engineering, or agentic workflows. Backend first, AI second.

The pipeline they describe breaks into four rentable layers:

Layer	What it does	What the listing points at
Transport	Carries audio in real time, handles interruptions	LiveKit
ASR / STT	Speech to text	Optimising this is explicitly in the role
Reasoning	Decides what to say next	Prompting strategy is owned by this role
TTS	Text back to speech	Latency work called out

Three of those four are API calls. The engineering is in gluing them together so a person on the other end doesn't hear a two-second gap before every reply. The listing also asks for testing frameworks for agents, which is the part most people skip and the part that actually determines whether the thing ships.

If you have written Node services and touched AWS, you're closer to this role's requirements than to a frontend job you'd apply for out of habit.

---

🇱🇰 The Sinhala and Tamil gap is the real opening

Here's the part worth your attention. Every call centre in Colombo doing collections, appointment reminders, insurance follow-ups, or delivery confirmations is doing exactly what CollectWise automated, except in Sinhala and Tamil. Nobody has automated it well.

The constraint is real and I won't hand-wave it. The support picture splits cleanly:

• Speech models handle Sinhala and Tamil. Whisper lists si and ta among its transcription languages. Gemini's TTS voice list includes both.
• Translation models handle them. NLLB-200 covers sin_Sinh and tam_Taml.
• Flagship chat models do not officially list them. Neither Sinhala nor Tamil appears in the published benchmark tables for GPT, Claude, or Llama's stated language sets.

So the ears and the mouth exist. The brain is the uncertain layer, and it's uncertain in a way the vendors haven't documented for you. Before you commit to a model for a Sinhala voice agent, check what it actually claims to support with our AI model language support checker, then test it on your own transcripts anyway. Documentation is a starting point, not evidence.

Code-switching is the harder problem nobody's benchmark measures. Real Sri Lankan phone calls run Sinhala grammar with English nouns in the same sentence. Test on that, not on clean Sinhala.

---

💰 Why the economics work here even better

The reason a voice agent business is attractive in the US is labour cost. That logic is weaker in Sri Lanka, where a call centre agent costs a fraction of a US collector. So does the opportunity vanish?

No, it moves. Three things still hold locally:

1. Coverage, not cost. An agent works at 9pm, on Poya day, and during the third call attempt nobody wants to make. Humans quietly stop dialling; software doesn't.
2. Consistency. Every call follows the script, logs cleanly, and produces structured data. Most local operations can't tell you what was said on call 400.
3. Language reach. One system serving Sinhala, Tamil, and English callers without staffing three teams.

The business case is throughput and record-keeping, not payroll savings. Pitch it that way and it survives the first meeting.

---

⚖️ One warning before you build a collections bot

CollectWise is in debt collection, which is a regulated activity almost everywhere. Automated calls that pressure people about money sit close to consumer protection rules, and I'm not going to pretend I know the current Sri Lankan position well enough to advise you on it.

Check the regulatory position with a lawyer before you automate a single collections call. Recorded-call consent, contact frequency, and disclosure requirements are the ones that bite.

If that's too much friction, the same stack applies to appointment reminders, order confirmations, clinic bookings, and delivery scheduling. Same technology, far less legal surface.

---

🛠️ What this means for you

The job is closed to you. The pattern isn't.

• This week: build one voice loop. Mic in, transcript out, LLM response, audio back. Do it in English first and measure end-to-end latency. If it's above one second, you have your engineering problem.
• Next: swap in Sinhala. Record twenty real-sounding utterances, including code-switched ones, and measure transcription accuracy yourself. That number is your actual product risk.
• Then: pick one narrow, low-stakes use case with a local business. Appointment reminders for a dental clinic beats a general-purpose assistant every time.
• Skill-wise: if you're a backend engineer, you're already most of the way to what this listing asks for. Add real-time audio and agent evaluation, and you're the profile a company like this pays $200K+ for.

The signal in this job ad isn't that AI can talk. It's that eight people built a real business on it in under two years, and the same stack is sitting unused for the two languages most of this country actually takes phone calls in.