Frontier labs post-train on the open internet

We post-train on data the internet has never seen handwritten deeds in six languages, regional-dialect voice conversations, survey maps and satellite imagery, decades of skewed scans

Off-the-shelf models, even the best ones, fail on this distribution

Closing that gap is the job

Terra is a multi-agent system where models converse with users in voice and text, read documents and site photos, reason over video and GIS data, and make judgment calls that lenders and families act on

Youll own post-training across this surface: adapting open-weight models language, vision-language, speech, and beyond to the modalities and behaviors Terra needs, end to end from data curation through training, evals, and deployment

This is hands-on training work in production, not prompt engineering with extra steps

What Youll Work On

- Multimodal post-training

SFT and preference optimization (DPO/GRPO-class methods) on open-weight models across text, vision, and speech document understanding today; voice agents, video, and geospatial reasoning on the near-term roadmap

- Agentic behavior tuning

Training models to be valuable

agents

, not just good predictors tool use, multi-step reasoning, code-switching across Indian languages, knowing when to escalate versus answer, and grounding every claim in evidence

- Provenance-grounded outputs

Structured generation where answers carry their evidence schema design, constrained decoding, and training strategies that make models cite rather than hallucinate

In our domain, a confident wrong answer costs someone their home

- Speech and voice adaptation

Adapting ASR/TTS and speech-LLM models for Indian languages, accents, and the messy acoustics of real phone calls so Terra can serve users who will never type

- GIS and visual grounding

Teaching models to reason over survey maps, plot boundaries, and satellite imagery, and to reconcile them with what documents claim

- Training data as a product

Annotation taxonomy design, mining production transcripts and failures into training data, and building the flywheel that compounds quality across every modality

- Evals that gate releases

Evaluation suites per modality and per agent behavior, with regional quality floors a model that works in Telangana but breaks in Karnataka never ships

What Were Looking For

- 3 7 years in ML engineering with hands-on post-training experience:



youve personally run SFT or preference-optimization jobs on open-weight models and shipped the result, not just read the papers

- Depth in at least one modality beyond text vision-language, speech, video, or geospatial and the appetite to expand into the others

Nobody arrives knowing all of them; we care that youve gone deep once and can do it again

- Fluency with the modern training stack: PyTorch, Hugging Face ecosystem (transformers, PEFT, TRL or similar), experiment tracking, GPU training workflows

- Strong evaluation instincts you design the eval before the training run, and you explain regressions from error analysis, not vibes

- Data-centric mindset: you know most post-training wins come from better data, and you have the tooling sense to build curation pipelines

- Solid Python and the engineering discipline to make training reproducible

Nice to Have

- Experience tuning models for agentic behavior: tool use, function calling, multi-turn dialogue, or RL on agent trajectories

- VLM fine-tuning (Qwen, Nemotron, Gemma class) or speech-model adaptation (Whisper-class ASR, speech LLMs, TTS) for Indic languages

- DPO/RLHF/GRPO experience, reward modeling, or preference-data design

- Inference optimization: vLLM, quantization, multi-adapter serving

- Remote sensing, GIS, or video understanding background

- Domains where provenance and correctness are non-negotiable legal, fintech, healthcare

Your First 90 Days

Days 1 30: Ground truth

Ship an improvement to a production model in your first two weeks

Read documents, listen to call recordings, study failure transcripts until you understand why this data breaks pretrained models

Own quality for one model surface end to end

Days 31 60: Own a modality

Take full ownership of one post-training track document VLMs, voice, or agent behavior including its data pipeline, eval suite, and a measurable quality lift shipped to production

Days 61 90: Shape the stack

Make the call on a structural bet preference optimization rollout, a new modalitys training approach, adapter consolidation backed by evidence from your first 60 days





By now you should be setting post-training direction across Terra, not just executing it

Why This Role

- Post-training scope that frontier labs split across whole teams text, vision, speech, and agents, all yours to shape

- A data distribution nobody else has: your work cant be replicated by anyone scraping the internet

- Direct stakes Terras judgments decide whether property purchases are safe, with lenders and families relying on the output

- Small, senior team; backed by Y Combinator and top investors, with real revenue and real users

About the interview

We move fast the full loop takes 5 7 days, and well give you a decision within 48 hours of your final round

- Intro call (30 min)

Mutual fit, your background, and a walkthrough of the problem space

- Post-training deep-dive (60 min)

A fine-tuning project you ran end to end: data decisions, training setup, what the evals caught, what youd redo

Were probing for hands-on depth

- Work sample (take-home or paired, your choice, ~3 hrs)

Design a post-training and eval approach for a real Terra problem in your strongest modality data strategy, training plan, and how youd know it worked

- Systems + eval round (60 min)

Design the full post-training loop for a multi-agent, multimodal system: curation training eval gates serving

Expect pushback

- Founder conversation (45 min)

Values, ambition, and your questions about where the company is going

About Landeed

Two thirds of Indian court cases are land related

Our solution to this is the Landeed, Indias fastest and most comprehensive title search engine

We are now actively growing our engineering and product teams to expand our title coverage to more Indian states and build an enterprise platform for government bodies and corporates alike

Founded: 2022

Batch: S22

Team Size: 35

Status:

Active

Location: Hyderabad, India

Founders

ZJ Lin

Founder

ZJ Lin

Founder

Sanjay Mandava

Founder

Sanjay Mandava

Founder

J Richards

Founder

J Richards

Founder

Similar Jobs

Jarmin

Fort

Ultra

DeepSim, Inc

Infer

Morph

Partcl

Seeing Systems

Pasito

Maywood

CellType

Knowtex

RentFlow

Trata

Hestus, Inc

Disclaimer: This job posting has been aggregated from external source. Role details, content, and availability are subject to change. Applicants are advised to confirm the latest information directly on the company website before applying.

📌 Training Engineer (Telangana)
🏢 Y Combinator
📍 Telangana