The Token Company — ML Researcher

Type: Full-time | On-site | San Francisco, CA Compensation: $150,000–$300,000 + 0.5%–1% equity Hiring count: 1 Visa sponsorship: Yes — H-1B, O-1, OPT Reports to: Founder

About The Token Company

The Token Company does LLM interpretability and context-optimization research, building custom machine learning models that analyze and compress token contexts before they reach the underlying model. The result is roughly 50% inference cost reduction, lower latency, and measurably higher accuracy for the enterprises and scale-ups integrating LLMs into their products. Seven months old with roughly 1,000 customers, the company raised $11.7M from First Round Capital and Y Combinator, with additional backing from the founders of Hugging Face, Slack, and Dropbox, and has been through both YC and HF0.

Founded: 2025 | Team size: 1–10 (Seed) | Total funding: $11.7M Industry: AI Tools Website: https://thetokencompany.com Office: San Francisco, CA

Why Candidates Should Join

• Research that ships: Success is measured by getting a model into a product used by ~1,000 customers, not by publications. You see your work in production quickly.
• Own a frontier problem end-to-end: Full ownership of a slice of LLM context compression and mechanistic interpretability — hypothesis, data, architecture, training, evals, and production impact.
• High autonomy: Every researcher directs their own agenda with minimal structure, reporting to the founder.
• Serious backing & pedigree: $11.7M from First Round Capital and Y Combinator, plus the founders of Hugging Face, Slack, and Dropbox; YC and HF0 alumni.
• Real compute: Training runs on NVIDIA B200s and large-scale GPU clusters.
• Everything covered: SF housing, food and meals, laundry and cleaning, healthcare and dental, significant equity, visa sponsorship, resources to build out the research team, and company off-sites.

Intake Call Summary

• No intake call transcript was available on the role page. An Intake Video is posted on Contrario but was not transcribed here — review it directly for hiring-manager nuance before scoring borderline candidates.

The Role

As an ML Researcher, you own a slice of one of the most interesting open problems in applied AI: figuring out what information inside an LLM context actually matters, and how to represent it more efficiently. This is a high-autonomy, high-output role for someone who wants to run a large volume of experiments, reproduce papers, and see their research ship into a production system used by real customers.

What You'll Be Doing

• Design and run experiments on LLM context compression and mechanistic interpretability, including model training, data curation, labeling pipelines, and evals
• Read current research papers and generate longer-term ideas for representing context more efficiently for LLMs
• Own your research direction end-to-end, from hypothesis through training runs on NVIDIA B200s and large-scale GPU clusters to evaluation and production impact
• Contribute to the eval infrastructure that measures how model outputs change and how compression affects accuracy and latency
• Iterate quickly on new architectures and training methods, treating shipping a model into the product as the primary success condition

Tech stack: Transformers, custom model training loops (data + architecture + training + evals), NVIDIA B200s and large-scale GPU clusters, eval infrastructure.

Requirements

• Prioritize production impact over publication metrics
• Own model training stack including data, architecture, training, evaluation, and shipping
• Trained models from scratch, end-to-end ownership of data, architecture, and training loop
• Strong ML fundamentals: transformers, mechanistic interpretability, LLM research
• High-agency researcher: self-directed, experiment-driven, not RAG or chatbot-only
• Spiky profile: exceptional pre-career achievement in competitions, research, or founding
• SF in-person, 996 intensity, hacker-house environment

Green Flags

• Pretrained a transformer model
• Serious post-training or RL experience on transformers
• Built novel architecture or training method with results
• Shipped trained models into production systems
• Experience in research labs, startups, or scale-ups
• Exceptional early-career achievement

Red Flags

• Experience mostly in RAG, agents, or prompt engineering
• Primary focus on fine-tuning existing models through APIs
• Preference for publishing papers over shipping models
• Work-life balance as a stated priority

Role Details

• Salary — $150,000–$300,000 (above $300K possible for exceptional candidates)
• Equity — 0.5%–1%
• Experience — 2+ years
• On-site policy — On-site in San Francisco; hacker-house environment; ~996 pace (9am–9pm, six days/week) or more
• Visa sponsorship — H-1B, O-1, OPT
• Employment type — Full-time
• Location — San Francisco, CA

Screening Questions

1. Phone number
2. Are you allowed to work in the United States?
3. LinkedIn / Personal website

Note: The Required Candidate Q&A section on the role page was collapsed ("Show More") — additional screening questions may exist beyond these three. Confirm the full list in Contrario before submitting.

Interview Process

Stage 1 — Pending Approval — Candidates awaiting initial approval. Stage 2 — Otso screen — Initial screen. Stage 3 — Second person technical — Technical interview with a second team member. Stage 4 — Whole team — Interview with the whole team. Stage 5 — Take home / Work trial — Practical work trial. Stage 6 — Offer — Offer extended. Stage 7 — Hired — Candidate accepts and starts.

Ideal Companies & Backgrounds

Updated July 16, 2026 — drawn from the role's Nice-to-Have list; no separate Contrario "Ideal Companies" section was present on the page.

University AI labs / frontier labs — Stanford AI Lab (SAIL), Berkeley AI Research (BAIR), or a frontier AI lab Competitive / exceptional pre-career achievement — Kaggle Grandmaster, IOI or ISEF medalist, ICPC finalist, or similar Early-stage experience — Prior startup experience, founding experience, or a technical lead role at an early-stage company Technical pedigree — Applied math, CS, or engineering background from a university with strong technical pedigree Model work — Has pretrained a transformer, done serious post-training or RL on one, or built a novel architecture with results