Lead AI Engineer

Vahan; Bengaluru; Sur place; INR 3 000 000 - 5 500 000; Plein temps

Résumé du poste

United States Digital Space LLC is seeking an AI Engineer III to own outcomes across the voice and WhatsApp agent systems, the eval suite, and data-driven improvements. You’ll diagnose metric shifts, justify fixes with leverage, and drive end-to-end impact in production systems for high-scale placement tasks.

You’ll work with async Python, evaluation pipelines, latency and cost optimization, and code-switching voice processing.

Qualifications

• Shipped an LLM-backed system in production.
• Can trace a business metric to its root cause across any layer.
• Strong async Python — you write services, not notebooks.
• Brings up evaluation before we do.
• Model-choice opinions grounded in latency and cost.
• Fluent with coding agents; can name the last time one produced something wrong.
• Can explain a technical trade-off to a recruiter or ops lead without dumbing it down.

Responsabilités

• Own the outcome across the voice and WhatsApp agents, the eval suite, and the conversion funnel.
• Write your own queries to understand funnel stages and metrics.
• Own agent behaviour end to end: loop, context, stop conditions, retries, and timeouts.
• Build and defend the eval suite with gold standards and regression gates.
• Address latency and cost; optimize time-to-first-audio and 100K calls/day.
• Develop the voice pipeline: code-switch Hinglish, turn detection, and network failure handling.
• Debug production nondeterminism; reproduce rare failures and write postmortems.
• Leverage coding agents (Claude Code) from Day 1.

Connaissances

LLM in production

Root-cause analysis

Async Python

Evaluation mindset

Model selection

Coding agents

Explain trade-offs

Learning in public

Outils

LiveKit

Pipecat

Deepgram

ElevenLabs

Cartesia

Description du poste

Role title: AI Engineer III

Location: Bengaluru (On-site)

Experience: 5+ years

About the company

the company is a marketplace for blue-collar work. Zomato, Zepto, Swiggy and others hire through us; hundreds of recruitment agencies source through us. Samvaadini, our hybrid voice and WhatsApp agent, places 100,000+ outbound calls a day to people looking for delivery and warehouse work across 900+ Indian cities — qualifying them, answering questions about pay and shift timings, and handing the serious ones to a recruiter. It speaks Hindi, Hinglish, and a widening set of regional languages, mostly to people on low-end Android phones in noisy streets, many of whom have never used a chatbot before. When Samvaadini gets better, more people get placed in a job this week instead of next month.

The Role

You’ll own outcomes across the system, not one component — the voice and WhatsApp agents, the eval suite, the conversion funnel they feed, and root-cause analysis across code, agent behaviour, models, and data. At this level you’ll more often be handed a number that moved than a task to complete: qualified handoffs down 12%? You localize it across code, agent behaviour, a silent model change, upstream data, or a market shift, then argue for the fix with the most leverage.

What You’ll Do

• Own the outcome, not the ticket — investigate metric moves end to end and argue for the highest-leverage fix (sometimes that’s doing nothing)
• Know the funnel cold: outreach, connect, qualify, handoff, interview, placement — write your own queries
• Own agent behaviour end to end: the loop, tool surface, context management, stop conditions, retries, and timeout handling
• Build and defend the eval suite — every behaviour change ships behind a golden set and a regression gate
• Fight for latency and cost — time-to-first-audio and inference cost at 100K calls/day are yours to defend
• Work the voice pipeline — streaming ASR on code-switched Hinglish, turn detection, barge-in, graceful failure on network drops
• Debug production non-determinism — reproduce rare failures, write postmortems, close the loop with evals
• Use coding agents hard and well — Claude Code from Day 1

What We’re Looking For

• Shipped an LLM-backed system in production, with real users, long enough for it to break
• Can trace a business metric to its root cause across any layer — as comfortable in a warehouse query as a stack trace
• Strong async Python — you write services, not notebooks
• Brings up evaluation before we do
• Model-choice opinions grounded in latency and cost — and you change them when shown better numbers
• Fluent with coding agents; can name the last time one confidently produced something wrong and how you caught it
• Can explain a technical trade-off to a recruiter or ops lead without dumbing it down
• Learns fast in public

Nice to Have (Genuinely Not Required)

Real-time voice stacks (LiveKit, Pipecat, Deepgram, ElevenLabs, Cartesia, telephony); Hindi/Hinglish or code-switched NLP; fine-tuning/distillation; durable workflow orchestration; ClickHouse or comparable analytical SQL; open-source work or public writing; MCP server authoring.