# AI-Startup Interview Difficulty Index (2026)

Source: https://www.techinterview.org/ai-startup-interview-difficulty-index/
Updated: 2026-09-08 · techinterview.org

**TL;DR —** We scored the engineering interviews at 74 AI-native companies on a five-part rubric — coding, system design, ML/research depth, take-home load, and speed/ambiguity — by reading each company's own interview guide and rating it 1 to 5 on every dimension. The frontier labs sit at the top (xAI 4.4, OpenAI 4.2), while applied-AI companies such as Oxide Computer and Yugabyte are the most approachable. Every score links to the full guide it came from. These are our editorial ratings drawn from each documented process, not survey data.Interview difficulty at an AI company is not one number, it is a shape: how hard the coding rounds get, whether there is a real system-design bar, how much frontier ML or research knowledge you need, whether a take-home is waiting, and how much the loop rewards speed and comfort with ambiguity. This index rates all five for every AI-native company we have a full interview guide for, so you can see at a glance where a company is genuinely hard and where the reputation runs ahead of the process.

## The index
Ranked hardest to easiest by overall difficulty. Higher is harder: 1 is a light bar, 5 is among the hardest anywhere. Overall is the plain average of the five columns.

| # | Company | Overall | Coding | System design | ML / research depth | Take-home | Speed & ambiguity |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | xAI | 4.4 | 5 | 4 | 5 | 3 | 5 |
| 2 | OpenAI | 4.2 | 4 | 4 | 5 | 3 | 5 |
| 3 | Cursor | 4.0 | 4 | 3 | 4 | 4 | 5 |
| 4 | Fireworks AI | 4.0 | 4 | 3 | 5 | 4 | 4 |
| 5 | Magic.dev | 4.0 | 3 | 4 | 5 | 4 | 4 |
| 6 | Mistral AI | 3.8 | 5 | 2 | 5 | 4 | 3 |
| 7 | Reflection AI | 3.8 | 4 | 4 | 5 | 2 | 4 |
| 8 | Together AI | 3.8 | 4 | 3 | 5 | 4 | 3 |
| 9 | Tilde | 3.8 | 3 | 4 | 5 | 3 | 4 |
| 10 | Reka AI | 3.6 | 4 | 4 | 5 | 1 | 4 |
| 11 | Skild AI | 3.6 | 4 | 4 | 5 | 1 | 4 |
| 12 | Scale AI | 3.6 | 3 | 3 | 4 | 4 | 4 |
| 13 | Exa | 3.6 | 2 | 4 | 3 | 5 | 4 |
| 14 | OpenEvidence | 3.6 | 2 | 4 | 3 | 5 | 4 |
| 15 | Modal | 3.4 | 4 | 4 | 2 | 4 | 3 |
| 16 | Black Forest Labs | 3.4 | 4 | 4 | 5 | 1 | 3 |
| 17 | ElevenLabs | 3.4 | 4 | 4 | 4 | 1 | 4 |
| 18 | Pika Labs | 3.4 | 4 | 4 | 4 | 1 | 4 |
| 19 | Parallel | 3.4 | 4 | 4 | 3 | 2 | 4 |
| 20 | Anthropic | 3.4 | 4 | 3 | 4 | 3 | 3 |
| 21 | Sierra | 3.4 | 3 | 4 | 3 | 3 | 4 |
| 22 | Stability AI | 3.4 | 3 | 4 | 5 | 2 | 3 |
| 23 | Cartesia | 3.4 | 3 | 4 | 5 | 1 | 4 |
| 24 | fal.ai | 3.4 | 3 | 4 | 4 | 2 | 4 |
| 25 | Legora | 3.4 | 3 | 4 | 3 | 3 | 4 |
| 26 | Ambience Healthcare | 3.4 | 2 | 4 | 3 | 4 | 4 |
| 27 | Glean | 3.2 | 4 | 4 | 3 | 1 | 4 |
| 28 | Suno | 3.2 | 4 | 3 | 4 | 1 | 4 |
| 29 | Contextual AI | 3.2 | 4 | 3 | 4 | 1 | 4 |
| 30 | Decagon | 3.2 | 3 | 4 | 4 | 1 | 4 |
| 31 | Deepgram | 3.2 | 3 | 4 | 4 | 1 | 4 |
| 32 | Replit | 3.2 | 3 | 4 | 3 | 2 | 4 |
| 33 | The Browser Company | 3.2 | 3 | 4 | 2 | 3 | 4 |
| 34 | Factory AI | 3.2 | 3 | 4 | 3 | 2 | 4 |
| 35 | Twelve Labs | 3.2 | 3 | 4 | 4 | 1 | 4 |
| 36 | Commure | 3.2 | 3 | 4 | 3 | 2 | 4 |
| 37 | LangChain | 3.2 | 3 | 3 | 3 | 3 | 4 |
| 38 | Perplexity | 3.2 | 3 | 3 | 4 | 2 | 4 |
| 39 | Physical Intelligence | 3.2 | 3 | 3 | 5 | 1 | 4 |
| 40 | Luma AI | 3.2 | 3 | 3 | 4 | 2 | 4 |
| 41 | Prime Intellect | 3.2 | 3 | 3 | 5 | 1 | 4 |
| 42 | Assort Health | 3.2 | 3 | 3 | 2 | 4 | 4 |
| 43 | Poolside | 3.2 | 2 | 4 | 5 | 1 | 4 |
| 44 | MotherDuck | 3.0 | 4 | 3 | 2 | 4 | 2 |
| 45 | LiveKit | 3.0 | 3 | 4 | 2 | 2 | 4 |
| 46 | XBOW | 3.0 | 3 | 3 | 4 | 1 | 4 |
| 47 | Turbopuffer | 3.0 | 2 | 4 | 2 | 2 | 5 |
| 48 | Rillet | 3.0 | 2 | 4 | 1 | 4 | 4 |
| 49 | Infinitus Systems | 3.0 | 2 | 4 | 3 | 2 | 4 |
| 50 | CoreWeave | 2.8 | 3 | 4 | 2 | 1 | 4 |
| 51 | Character.AI | 2.8 | 3 | 3 | 4 | 1 | 3 |
| 52 | Hugging Face | 2.8 | 3 | 3 | 3 | 3 | 2 |
| 53 | Runway | 2.8 | 3 | 3 | 4 | 1 | 3 |
| 54 | Sakana AI | 2.8 | 3 | 3 | 5 | 1 | 2 |
| 55 | Abridge | 2.8 | 3 | 3 | 4 | 1 | 3 |
| 56 | Genspark | 2.8 | 3 | 3 | 3 | 1 | 4 |
| 57 | Saronic | 2.8 | 3 | 3 | 3 | 1 | 4 |
| 58 | Shield AI | 2.8 | 3 | 3 | 3 | 2 | 3 |
| 59 | OpenRouter | 2.8 | 2 | 4 | 2 | 2 | 4 |
| 60 | Tennr | 2.8 | 2 | 3 | 2 | 4 | 3 |
| 61 | Baseten | 2.6 | 3 | 4 | 2 | 1 | 3 |
| 62 | Tempus AI | 2.6 | 3 | 3 | 3 | 1 | 3 |
| 63 | Mercor | 2.6 | 3 | 3 | 2 | 1 | 4 |
| 64 | Rogo | 2.6 | 3 | 3 | 3 | 1 | 3 |
| 65 | Dagster | 2.6 | 3 | 3 | 1 | 3 | 3 |
| 66 | Cohere | 2.6 | 3 | 2 | 3 | 3 | 2 |
| 67 | Harvey | 2.6 | 2 | 3 | 3 | 1 | 4 |
| 68 | Taktile | 2.6 | 2 | 3 | 2 | 3 | 3 |
| 69 | Hippocratic AI | 2.4 | 3 | 3 | 3 | 1 | 2 |
| 70 | Crusoe | 2.4 | 2 | 3 | 2 | 1 | 4 |
| 71 | TensorWave | 2.4 | 2 | 3 | 2 | 1 | 4 |
| 72 | Mach Industries | 2.2 | 2 | 2 | 1 | 1 | 5 |
| 73 | Yugabyte | 2.0 | 3 | 3 | 1 | 1 | 2 |
| 74 | Oxide Computer | 2.0 | 1 | 2 | 1 | 4 | 2 |

## How we scored it
For each company we read its full techinterview.org interview guide and rated the five dimensions from 1 to 5, using only what the guide documents about the process. The overall score is the equal-weight average of the five, so a company that is brutal on coding but has no take-home can land in the same place as one that is moderate everywhere. Ties are broken by the coding score, then system design.

| Dimension | What it measures | A 1 looks like | A 5 looks like |
| --- | --- | --- | --- |
| Coding | How algorithmic and demanding the coding rounds are. | Practical, working-code questions, no puzzles. | Multiple hard rounds at a top-lab bar, implemented from scratch. |
| System design | Depth and number of system- or ML-system-design rounds. | No dedicated design round. | Several deep design rounds on real infra or model-serving. |
| ML / research depth | How much frontier ML or research knowledge is expected. | Applied or product engineering, little ML needed. | Paper-level model internals and training/inference math. |
| Take-home | Presence and weight of a take-home or work sample. | None. | A large, heavily weighted take-home or paid work trial. |
| Speed & ambiguity | How much the loop signals hypergrowth pace and ownership. | Measured, structured, and calm. | Fast loops, high ambiguity, strong culture and bar-raiser filtering. |
methodologyThese are our editorial ratings of each company's *documented* interview process, applied on one fixed rubric so companies can be compared on the same terms. They are not a survey, and they are not measured statistics. Loops change, levels differ, and any single candidate's experience will vary, so treat a score as a starting read and check it against the linked guide before you plan around it.
## What the index shows
The hardest interviews belong to the frontier labs. xAI, OpenAI, Cursor and Fireworks AI cluster at the top because they stack three things at once: a top-tier coding bar, real research and model-internals depth, and loops built for speed and intensity. If you are aiming there, the coding grind is the floor, not the ceiling, and the ML depth is what actually separates offers from rejections.

The most approachable loops, Oxide Computer, Yugabyte, Mach Industries and TensorWave among them, are the applied-AI companies building products on top of models rather than training their own. The coding tends to be LeetCode-medium and practical, the design round is about agents, retrieval, or serving rather than GPU-scale infrastructure, and deep research knowledge is optional for most roles. They are still real bars, but the shape rewards strong product engineers over researchers.

Two patterns hold across the list. Take-homes are the exception, not the rule: most of these companies run live rounds and reserve a longer work sample for senior or research hires. And ML/research depth is the single dimension that most reliably tracks a company's overall difficulty, which is why an infrastructure company like Modal can score high on coding and design yet land in the middle overall.

## Scoring notes, company by company
The one-line rationale behind each rating, so you can see exactly what drove it.

- [**xAI**](https://www.techinterview.org/companies/xai/) 4.4 — Frontier-lab coding/ML bar like OpenAI; one deep GPU-scale design round; 4-8h take-home for some roles; Musk-pace intensity filters hard
- [**OpenAI**](https://www.techinterview.org/companies/openai/) 4.2 — Coding at Meta-E5/Google-L5 bar; 1 deep design round w/ inference math; world-class research rounds; role-dependent 4-8h take-home; loops prize shipping speed
- [**Cursor**](https://www.techinterview.org/companies/cursor/) 4.0 — Two medium-hard/hard coding rounds; 1 design round; frontier-adjacent ML deep-dive for ML roles; substantial 4-8h take-home; intense hypergrowth pace and select
- [**Fireworks AI**](https://www.techinterview.org/companies/fireworks-ai/) 4.0 — Kernel coding ~Nvidia/frontier bar; one design round; deep CUDA/FlashAttention/quant ML rounds; 4-8h take-home for senior; fast perf-obsessed culture
- [**Magic.dev**](https://www.techinterview.org/companies/magic-dev-interview-guide/) 4.0 — Frontier long-context lab; short fast 1-3wk loop, high ML bar; 5hr take-home is main filter; design on distributed training/inference
- [**Mistral AI**](https://www.techinterview.org/companies/mistral/) 3.8 — Research-eng coding is above LeetCode-hard (impl attention/RoPE from scratch); deep MoE/transformer internals expected; only 1 design round; 4-8hr take-home; fa
- [**Reflection AI**](https://www.techinterview.org/companies/reflection-ai-interview-guide/) 3.8 — High coding bar all tracks, implement PPO from scratch, ML-system design, RL near-mandatory for research, no take-home, fast small-team loop.
- [**Together AI**](https://www.techinterview.org/companies/together-ai/) 3.8 — Two 'solidly hard' CUDA/ML-systems coding rounds; deep inference-optimization/research rounds; 4-8hr take-home for senior/research; one design round; pace 'fast
- [**Tilde**](https://www.techinterview.org/companies/tilde-interview-guide/) 3.8 — Founder-led loop: deep GPU/Triton kernel and training-infra design, interpretability/SAE research taste; coding is practical, possible work trial.
- [**Reka AI**](https://www.techinterview.org/companies/reka-ai-interview-guide/) 3.6 — Frontier multimodal lab; deep ML research bar, PyTorch medium-hard coding, one ML-design round, no take-home, fast ownership-driven culture.
- [**Skild AI**](https://www.techinterview.org/companies/skild-ai-interview-guide/) 3.6 — Strong timed coding + ML system design; deep robot-learning/RL/sim-to-real research bar (PhD, top venues); no take-home; frontier-lab bar, hypergrowth $14B star
- [**Scale AI**](https://www.techinterview.org/companies/scale-ai/) 3.6 — Medium-hard coding ~mid-FAANG; single ML-data design round; real RLHF/eval depth for ML roles; weighted 4-6hr take-home; high ambiguity/pace
- [**Exa**](https://www.techinterview.org/companies/exa-interview-guide/) 3.6 — Light 15-min algo screen; weighty take-home + paid trial day; deep search/retrieval design; applied retrieval ML not frontier; fast build-quickly loop
- [**OpenEvidence**](https://www.techinterview.org/companies/openevidence-interview-guide/) 3.6 — CTO weights judgment/autonomy over coding drills; heavy 1-2wk take-home is the loop's defining feature; RAG/retrieval design-heavy; on-site, no-PM ownership, sm
- [**Modal**](https://www.techinterview.org/companies/modal/) 3.4 — Medium-hard coding (Stripe/HashiCorp bar), rigorous systems design + infra deep-dive; GPU-infra not ML research; substantive 4-8h take-home for senior; fast but
- [**Black Forest Labs**](https://www.techinterview.org/companies/black-forest-labs-interview-guide/) 3.4 — Frontier diffusion/flow-matching depth; ML-flavored medium-hard coding + ML sys design; no take-home; calmer culture but high ownership/taste.
- [**ElevenLabs**](https://www.techinterview.org/companies/elevenlabs-interview-guide/) 3.4 — 3 coding rounds medium-hard ML-flavored DSA; deep ML-system-design (TTS/codecs/CUDA); research deep-dive; no take-home; taste-driven fast-shipping startup
- [**Pika Labs**](https://www.techinterview.org/companies/pika-labs-interview-guide/) 3.4 — Multiple medium-hard coding/ML rounds plus dedicated ML system-design and video-diffusion depth; no take-home; small fast-shipping Series B startup stressing ow
- [**Parallel**](https://www.techinterview.org/companies/parallel-interview-guide/) 3.4 — Strong Python coding bar; deep retrieval/search design; ML mainly for model-training roles; optional take-home; hypergrowth in-person team.
- [**Anthropic**](https://www.techinterview.org/companies/anthropic/) 3.4 — Coding ~Google L5; one design round; ML/paper depth for research roles; role-dependent 4-8h take-home; careful, document-driven, values-filtered.
- [**Sierra**](https://www.techinterview.org/companies/sierra-interview-guide/) 3.4 — Practical TS/Py concurrency, not LeetCode; heavy agent system design; paid work sample; applied LLM not research; high-bar hypergrowth
- [**Stability AI**](https://www.techinterview.org/companies/stability-ai-interview-guide/) 3.4 — Diffusion/flow-matching depth + paper round; medium-hard ML-flavored coding; one ML sys-design round; ownership valued, calm post-reorg pace
- [**Cartesia**](https://www.techinterview.org/companies/cartesia-interview-guide/) 3.4 — Frontier SSM/Mamba research bar high; ML system-design on latency; coding real not LeetCode; no take-home; fast startup under ambiguity
- [**fal.ai**](https://www.techinterview.org/companies/fal-ai-interview-guide/) 3.4 — Systems-heavy bar; clean Python under time pressure over LeetCode; deep inference/GPU/CUDA design; optional short take-home; small fast in-person team values au
- [**Legora**](https://www.techinterview.org/companies/legora-interview-guide/) 3.4 — Medium coding under time; heavy LLM/retrieval system-design case with live 60-min presentation; applied LLM not research; fast-moving startup, founder final.
- [**Ambience Healthcare**](https://www.techinterview.org/companies/ambience-healthcare-interview-guide/) 3.4 — LRU-cache-level coding (Glassdoor ~2.6); deep EHR system design (rate limits, multi-tenancy); paid rubric take-home w/ presentation; mission-first urgency, nine
- [**Glean**](https://www.techinterview.org/companies/glean-interview-guide/) 3.2 — LeetCode mediums+some hards, clean-code bar; search/RAG design plus senior IR depth round; ML only for ML-track; no take-home; fast loop, high bar-raiser
- [**Suno**](https://www.techinterview.org/companies/suno-interview-guide/) 3.2 — 3 coding rounds medium-hard DSA; 1 ML sys-design; deep audio-ML/codecs + research deep-dive; no take-home; fast taste-driven startup
- [**Contextual AI**](https://www.techinterview.org/companies/contextual-ai-interview-guide/) 3.2 — Real live DS&A every track; 1-2 ML system-design rounds; deep RAG/retrieval focus; no take-home; small early team, ship under ambiguity
- [**Decagon**](https://www.techinterview.org/companies/decagon-interview-guide/) 3.2 — Practical TS/Python coding not puzzles; deep LLM system-design (agents/RAG/evals); applied AI fluency + papers; no take-home; fast frontier-startup bar.
- [**Deepgram**](https://www.techinterview.org/companies/deepgram-interview-guide/) 3.2 — Medium DSA coding; deep ASR/ML-systems design round plus ASR architecture knowledge; no take-home; fast-moving Series C startup stresses ownership/shipping
- [**Replit**](https://www.techinterview.org/companies/replit-interview-guide/) 3.2 — Practical working-code rounds (not leetcode); deep design on sandboxes/OT/agents; applied LLM not research; no big take-home; ship-it high-bar culture
- [**The Browser Company**](https://www.techinterview.org/companies/browser-company-interview-guide/) 3.2 — LeetCode-medium product-flavored coding; 3 design prompts incl AI-agent; 4-6hr take-home; small taste-driven team, high bar, ship-fast ownership; ML applied not
- [**Factory AI**](https://www.techinterview.org/companies/factory-ai-interview-guide/) 3.2 — Practical coding not LeetCode; heavy agent/LLM system-design; applied ML not research; no take-home; fast-growing funded startup, founder round
- [**Twelve Labs**](https://www.techinterview.org/companies/twelve-labs-interview-guide/) 3.2 — Medium coding, not the differentiator; deep video/multimodal ML + video system-design; no take-home; fast Series B ownership/speed bar.
- [**Commure**](https://www.techinterview.org/companies/commure-interview-guide/) 3.2 — Practical API/data-filtering screens over LeetCode; design weighted as the harder half; some ambient-ML eval/HIPAA questions; take-home only for non-eng; in-off
- [**LangChain**](https://www.techinterview.org/companies/langchain/) 3.2 — Two medium-hard coding rounds below Google L5; one LLM-platform design; applied-AI not frontier research; 4-6h take-home for senior/staff; fast pace + culture-f
- [**Perplexity**](https://www.techinterview.org/companies/perplexity/) 3.2 — Medium-hard coding below Google L5; one AI-search design round; distinctive retrieval/rerank/RAG deep-dive; take-home only some senior roles; fast ship-weekly p
- [**Physical Intelligence**](https://www.techinterview.org/companies/physical-intelligence-interview-guide/) 3.2 — Practical Python not LeetCode; ~1 ML system-design round; deep VLA/RL/flow-matching research bar; no take-home; small fast high-ownership 2024 startup
- [**Luma AI**](https://www.techinterview.org/companies/luma-ai-interview-guide/) 3.2 — Medium coding, not LeetCode-heavy; ML-system design for infra; deep diffusion/video research bar; notebook debugging not a take-home; CEO culture round, well-fu
- [**Prime Intellect**](https://www.techinterview.org/companies/prime-intellect-interview-guide/) 3.2 — Practical coding not LeetCode; one ML system-design round; deep distributed-training/RL research bar; no take-home; small fast startup screening hard for owners
- [**Assort Health**](https://www.techinterview.org/companies/assort-health-interview-guide/) 3.2 — Practical Python/FastAPI coding (rated 3.5/5), one real-time system-design round, week-long build-a-voice-agent take-home, hypergrowth 15->250 with founder roun
- [**Poolside**](https://www.techinterview.org/companies/poolside-interview-guide/) 3.2 — No LeetCode, just practical coding; deep RLCEF/training/eval/serving design; no take-home; short ~1wk loop, flat team screens for ownership
- [**MotherDuck**](https://www.techinterview.org/companies/motherduck/) 3.0 — Two medium-hard coding rounds + deep DuckDB internals; one design round; 4-8hr take-home common for senior; ML is applied text-to-SQL only; deliberate pace.
- [**LiveKit**](https://www.techinterview.org/companies/livekit-interview-guide/) 3.0 — Practical Go/Rust concurrency coding over LeetCode; deep distributed-systems + WebRTC design; occasional small take-home; founder round, post-Series C hypergrow
- [**XBOW**](https://www.techinterview.org/companies/xbow-interview-guide/) 3.0 — Practical coding over LeetCode; agent-platform design round; heavy LLM-agent+offensive-security depth; no take-home mentioned; tiny startup compresses loop to 1
- [**Turbopuffer**](https://www.techinterview.org/companies/turbopuffer-interview-guide/) 3.0 — No leetcode; deep systems/storage design, napkin math, perf debugging. Possible unverified work sample; reject-by-default P99 bar.
- [**Rillet**](https://www.techinterview.org/companies/rillet-interview-guide/) 3.0 — No LeetCode; practical coding. Weighted take-home plus ledger-flavored backend design. Non-ML product roles. Series B hybrid sprint pace.
- [**Infinitus Systems**](https://www.techinterview.org/companies/infinitus-interview-guide/) 3.0 — LC-medium coding (2.7/5 Glassdoor); hard part is long-lived voice-call system design; some LLM guardrail/eval depth; optional take-home; hires for agency/urgenc
- [**CoreWeave**](https://www.techinterview.org/companies/coreweave-interview-guide/) 2.8 — Medium DSA over 3 coding rounds; deep distributed/infra design + craft & networking deep-dives; infra not research ML; no take-home; heavy on-call/ownership, IP
- [**Character.AI**](https://www.techinterview.org/companies/character-ai-interview-guide/) 2.8 — LeetCode medium coding, 1 system design + 1 ML inference round; deep LLM serving knowledge expected; no take-home; 3-5wk startup-paced loop
- [**Hugging Face**](https://www.techinterview.org/companies/hugging-face-interview-guide/) 2.8 — Coding 'medium'/practical not LeetCode; 1-2 domain-depth rounds; ML pragmatism not frontier research; 2-4hr take-home some roles; explicitly measured pace, miss
- [**Runway**](https://www.techinterview.org/companies/runway-interview-guide/) 2.8 — Medium-hard DSA (2 coding rounds), one ML/product design round, video-diffusion depth + paper round for research; no take-home; ownership/taste focus.
- [**Sakana AI**](https://www.techinterview.org/companies/sakana-ai-interview-guide/) 2.8 — Research lab: heavy paper/ML-internals depth, medium-hard ML-flavored coding, 1 design round, no take-home, calm creativity-prized culture (not hypergrowth).
- [**Abridge**](https://www.techinterview.org/companies/abridge-interview-guide/) 2.8 — Timed DSA + one system-design round; deep ML/ASR eval probing for ML roles; no take-home; selective, deliberate month-long mission-fit loop.
- [**Genspark**](https://www.techinterview.org/companies/genspark-interview-guide/) 2.8 — Real DS&A coding bar; one system-design round on stateful jobs; applied agent/RAG focus not frontier research; no take-home; fast 1-3wk founder-led loop
- [**Saronic**](https://www.techinterview.org/companies/saronic-interview-guide/) 2.8 — Standard startup loop, ~3/5 difficulty, medium coding in C++/Python; domain depth in perception/autonomy; no take-home; fast defense-startup pace.
- [**Shield AI**](https://www.techinterview.org/companies/shield-ai-interview-guide/) 2.8 — LeetCode-medium C++ coding, domain-heavy (GNC/ML/real-time), optional ~4hr take-home for some teams, final presentation; Glassdoor 3.2/5.
- [**OpenRouter**](https://www.techinterview.org/companies/openrouter-interview-guide/) 2.8 — Practical coding over LeetCode; heavy API-gateway/routing/failover design; ML optional for platform roles; possible short take-home; lean startup, high-ownershi
- [**Tennr**](https://www.techinterview.org/companies/tennr-interview-guide/) 2.8 — Easy-to-medium LeetCode staples; one short rate-limiter/document-pipeline design; model is integrated not invented; heavily-weighted 48h API-to-API take-home; o
- [**Baseten**](https://www.techinterview.org/companies/baseten-interview-guide/) 2.6 — Practical infra coding not LeetCode; deep GPU-serving system design; ML-systems not research/papers; live rounds no take-home; selective but deliberate month-lo
- [**Tempus AI**](https://www.techinterview.org/companies/tempus-ai-interview-guide/) 2.6 — LeetCode-medium coding, one health-data design round, applied clinical ML + genomics depth for senior+, no take-home, committee-driven 4-6wk loop
- [**Mercor**](https://www.techinterview.org/companies/mercor-interview-guide/) 2.6 — LeetCode-medium coding with speed/clean-code bar; 1-2 rounds incl. system design; infra not research ML; live build not take-home; 6-day pace, fast 1-3wk loop
- [**Rogo**](https://www.techinterview.org/companies/rogo-interview-guide/) 2.6 — Medium coding bar; one design round for senior/staff; applied RAG/eval not frontier research; no take-home mentioned; short 2-4wk fast-growth loop
- [**Dagster**](https://www.techinterview.org/companies/dagster-interview-guide/) 2.6 — Practical Python + graph algorithms (topo sort), data-platform system design, occasional ETL take-home; small remote team, not ML research.
- [**Cohere**](https://www.techinterview.org/companies/cohere/) 2.6 — Two medium-hard coding rounds; one design round; ML/retrieval papers for research roles; 4-8h take-home for senior/research; fast but sustainable pace
- [**Harvey**](https://www.techinterview.org/companies/harvey-ai-interview-guide/) 2.6 — Coding is medium LeetCode, no exotic tricks; 1 LLM-app design round; applied LLM eval not model internals; no take-home; hypergrowth startup, ships fast, founde
- [**Taktile**](https://www.techinterview.org/companies/taktile-interview-guide/) 2.6 — Practical FastAPI take-home over LeetCode; one low-latency system-design round; ML-ops not research; organized 4-6wk loop w/ values round.
- [**Hippocratic AI**](https://www.techinterview.org/companies/hippocratic-ai-interview-guide/) 2.4 — Medium DSA coding; one ML system-design round on LLM safety/voice pipelines; applied LLM eng not research; no take-home; calmer-than-frontier pace.
- [**Crusoe**](https://www.techinterview.org/companies/crusoe-interview-guide/) 2.4 — Practical non-LeetCode coding (~2.8/5); one systems/design round on cloud control plane; infra not ML research; no take-home; screens hard for urgency/ownership
- [**TensorWave**](https://www.techinterview.org/companies/tensorwave-interview-guide/) 2.4 — Practical coding, explicitly not LeetCode-heavy; one infra-design round on GPU clusters; ROCm/ML-systems not frontier research; no take-home; lean fire-fighting
- [**Mach Industries**](https://www.techinterview.org/companies/mach-industries-interview-guide/) 2.2 — Hardware defense startup: project walkthroughs, first-principles physics/C++ estimation, no LeetCode-hard or take-home; intense fast in-person 10-12hr culture.
- [**Yugabyte**](https://www.techinterview.org/companies/yugabyte-interview-guide/) 2.0 — Medium DSA bar (3.2/5) with DP; deep concurrency, Raft/LSM/MVCC internals + one distributed KV design round. No take-home, no ML; ~2-3wk pace.
- [**Oxide Computer**](https://www.techinterview.org/companies/oxide-computer-interview-guide/) 2.0 — No leetcode, whiteboard, or live coding; hiring is written materials (work/writing/analysis samples) read before any talk. Systems depth, not ML. No hypergrowth

## Keep going
Every company above links to its full interview guide. For the complete set, see the [AI-native company interview guides](/ai-startup-interview-guides/) hub, which covers the frontier labs and the startups building on them.