AI Engineer

Help increase the GDP of the internet as part of our AI team About the role Flam builds multimodal AI systems that ship to real customers. Our stack spans four production surfaces: Falcon — Our LLM system, built on a sparse-Mo E backbone. Finesse — Our TTS and cross-lingual voice cloning system. We are a small team. Whoever joins will own systems end-to-end training, evaluation, serving, and the latency budget. RESPONSIBILITIES: Fine-tune and evaluate open-weight models Lo RA/QLo RA, full SFT, preference tuning) and get them into production without a quality regression. Optimize inference: quantization FP8/NVFP4/INT4, speculative decoding, prefix and KV-cache strategies, batching and scheduling on v LLM or SGLang. Build evaluation harnesses that tell us something true — benchmark suites, regression gates, and per-release comparisons against both our own prior checkpoints and external baselines. Own latency. Profile the pipeline, find where the milliseconds go, and remove them. Take research to production: read the paper, replicate it, decide honestly whether it's worth shipping, and then ship it. Write and maintain the serving infrastructure around your models —containers, autoscaling, GPU scheduling, observability What we're looking for Required 2+ years building ML systems that ran in production, not only in notebooks. Strong Python and Py Torch. You can read a model implementation and modify it, not just call .fit. Hands-on experience with at least one modern inference stack (v LLM, SGLang, Tensor RTLLM, or TGI) and a real understanding of what makes it fast. Demonstrated fine-tuning experience — you've trained something, evaluated it properly, and know why your eval numbers meant what you claimed. Comfort with GPU-level reasoning: memory layout, precision trade-offs, where the bottleneck actually is. Ability to work from a paper. We move on recent research and expect you to be able to read it. Strongly preferred Experience in one or more of: speech ASR/TTS, diffusion and image generation, or video/avatar generation. Quantization experience beyond running an off-the-shelf script. Real-time streaming systems Web RTC, low-latency audio/video pipelines). Indic language NLP, multilingual tokenizer work, or code-mixed data. Cloud GPU deployment on GCP, AWS, or Run Pod — including the unglamorous parts. Our stack Py Torch · v LLM · SGLang · Tensor RT / ONNX · Unsloth · LLa MAFactory · TRL Milvus · BGEM3 · Redis · NATS · Docker · GCP · Run Pod · Web RTC · Comfy UI.