Y Combinator · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 0:15

   The Hook

   Introducing SynSci: AI Co-Scientists for Research

   “I'm Ishaan and I'm Aayam from Synthetic to introduce you to SynSci, I'm going to ask it if we can fine-tune ESM-2 to predict protein mutation effects well enough to guide wet-lab experiments without deep mutational scanning. Benchmark on ProteinGym. I want to match or beat published results with minimal compute. Run the full workflow. From here, it's in the driver's seat.”

   On screen

   new session main branch last modified 1 minute ago Fine-tune ESM-2 to predict protein mutation effects well enough to guide wet-lab experiments without deep mutational scanning. Benchmark on ProteinGym. I want to match or beat published results with minimal compute. Run the full workflow. Research AI Claude Opus 4.6 Max ESM-2 LoRA Fine-Tuning on ProteinGym Considering next steps 2.1s status Research AI Claude Opus 4.6 Max

   Camera

   Static medium shot of two founders, then a close-up screen recording of the SynSci interface.

   Motion

   Direct cut from live-action to screen recording, with text highlighting and typing animation on the UI.

2. 0:16 – 0:47

   Product Reveal

   Automated Research Loop: Literature to Experiment

   “reads through dozens of papers, estimates compute costs, it spins up GPUs and launches the training job on Modal. Training finishes, we're up 10 points on Spearman.”

   On screen

   Gathering thoughts 47s skill Loaded skill citation-management skill Loaded skill research-lookup skill Loaded skill citation-management skill Loaded skill literature-review skill Planning next steps 8m, 52s Good. Modal credentials are live. Let me fetch the key papers for reference. Shell Download ProteinGym NeurIPS paper PDF Shell Download Schmeiser LoRA fine-tuning paper To-dos 3/8 literature review: ESM-2 and PLMs for mutation effect prediction literature review: ProteinGym benchmark and supervised methods literature review: LoRA/PEFT for protein language models Synthesize literature findings and define approach Implement ESM-2 fine-tuning pipeline with LoRA + regression head Deploying pipeline on Modal (disconnect-safe) Analyze results: compare zero-shot vs fine-tuned vs literature Download artifacts and present final report ask amy Research AI Claude Opus 4.6 Max Apps Logs Secrets Storage Notebooks esm2-proteingym-pipeline / App Logs Overview Deployment History App Logs Search functions run_pipeline Components: 1 live, 1 idle, 1 running Timestamp Feb 19 19:20:55.467 Epoch 1/1: 31% Feb 19 19:21:35.747 Epoch 1/1: 33% Feb 19 19:22:18.021 Epoch 1/1: 34% Feb 19 19:22:59.700 Epoch 1/1: 35% Feb 19 19:23:31.133 Epoch 1/1: 36% Feb 19 19:24:12.387 Epoch 1/1: 37% Feb 19 19:24:53.769 Epoch 1/1: 38% Feb 19 19:25:19.446 Epoch 1/1: 39% Feb 19 19:26:01.960 Epoch 1/1: 40% Feb 19 19:26:44.133 Epoch 1/1: 41% Feb 19 19:27:28.992 Epoch 1/1: 42% Add filter Function: run_pipeline Input ID: in-0XWKG778XW49XW074.EE X Function Call ID: fc-0XWKG778XW49XW074XQ4DXQ4MKS X Expand logs Search logs (3K) Settings Modal Modal 2028 About Status Changelog Documentation Slack Community Pricing Examples linkedin yep sure.. design it well though and make it local disconnect safe Hide steps 166m, 8.8s results.esm-2 850m fine-tuning for Mutation Effect Prediction Test Set (41 held-out assays) Method Mean Spearman Median p p ESM-2 850M zero-shot 0.431 0.476 ESM-2 850M fine-tuned 0.531 0.579 (ours) Improvement +0.100 +0.103 70.7% of test assays improved over zero-shot. vs Published Baselines ask anything... "what is the tech stack of this project?"

   Camera

   Screen recording with dynamic UI elements, zooming into specific sections of the interface.

   Motion

   Scrolling text, expanding UI elements, progress bars, and data visualization (bar chart) with numerical updates. Subtle background color shifts.

3. 0:48 – 1:21

   Feature Teaser

   Full AI Loop & Continuous Improvement

   “also writes the whole thing up for you, all verified and ready to go. But let's be honest. Watch. I just switched to flywheel mode. It orchestrates ML scientists to deploy that same with every result feeding back into retraining. full loop. Train, evaluate, deploy, iterate. Iterate continuously gets sharper with your data,”

   On screen

   DMS measurements. 1.1 Model Architecture Figure 1 provides a schematic of the end-to-end architecture and how a mutation is mapped to a scalar prediction. Model Architecture for Protein Mutation Effect Prediction (ESM-2 + LoRA) INPUT Protein Sequence ESM-2 (850M parameters) POSITION EXTRACTION (1280d) Hidden Representations LoRA (r=8) (2M parameters) CONCATENATION REGRESSION HEAD Layer 1: 1408 to 512 (GELU, Dropout) Layer 2: 512 to 128 (GELU, Dropout) Layer 3: 128 to 1 (GELU) MSE Loss Actual (predicted DMS score) Ground Truth A142V (wildtype to mutant V at position 142) Extract h (1280d) Combined Vector (1408 dim) MUT embedding Figure 1: Architecture schematic for supervised mutation effect prediction. A wildtype protein sequence is encoded by ESM-2 with LoRA adapters. For a mutation at position p (e.g., A142V), we extract the final-layer hidden state h, and concatenate it with learned embeddings for the wildtype and mutant residues (ewt, emut). A small MLP predicts a scalar fitness score, optimized with mean squared error (MSE) against per-assay z-scored DMS labels. all files methods_results.pdf _vuff_cache kbxcnbh esm2_mutation b _pycache_ results results.json train_meta.json _API_DEPLOYMENT_TIPS.md _check_analysis.py _ddp_train.py _model_app.py _retrain_from_feedback.py _serve_api.py _train_and_eval.py _trigger.py figures methods_results.pdf methods_results.tex references.bib sigmac-bridge.ipynb Parameter-Efficient FI... 5/6 95% Table 1: Comparison with published baselines on the ProteinGym mutation benchmark. Published values are zero-shot, evaluated on all 217 assays. Our fine-tuned result is supervised, evaluated on 41 (or 20 clean) held-out assays from 301 filtered assays. Direct comparison is not strictly valid due to differing evaluation protocols. Method Input Evaluation AIDO Protein RAG [Li et al., 2023d] 0.518 Structure + MSA ZS, 217 assays ProT5st [K-2014b] [Li et al., 2019b] 0.507 Structure + Seq. ZS, 217 assays DeepSeq [Margetis et al., 2022] 0.458 MSA + Seq. ZS, 217 assays TranceptEVE [Niu et al., 2023] 0.456 MSA + Sequence ZS, 217 assays ESM-2 650M [Lin et al., 2022] 0.414 Sequence ZS, 217 assays ESM-2 650M zero-shot (ours) 0.431 Sequence ZS, 41 test assays ESM-2 650M + LoRA (ours) 0.531 Supervised 41 test ESM-2 + LoRA (clean) 0.525 Sequence + DMS Supervised, 30 test on the 30 clean assays without protein overlap (p = 0.025), performance remains competitive. However, the critical distinction is that our model is supervised. It has access to labeled DMS data from 100 training assays, whereas the published baselines operate in a zero-shot setting. This makes (e.g., dms) The practical implication is that when a model has access to training data for proteins, parameter-efficient fine-tuning of a protein language model can yield substantial improve- ments in mutation effect prediction at minimal computational cost (2M trainable parameters, 134 minutes on 4 H100 GPUs). This positions the approach as a lightweight complement to zero- shot methods, particularly for models with many parameters, as mutations often have a status ESM-2 fine-tuning for protein mutation effects unp sure.. design it well though and make it local disconnect safe Hide steps 166m, 8.8s results.esm-2 850m fine-tuning for Mutation Effect Prediction Test Set (41 held-out assays) Method Mean Spearman Median p p ESM-2 850M zero-shot 0.431 0.476 ESM-2 850M fine-tuned 0.531 0.579 (ours) Improvement +0.100 +0.103 70.7% of test assays improved over zero-shot. vs Published Baselines ask anything... "what is the tech stack of this project?" Research GPT-3.2 ESM-2 Mutation Prediction API with Feedback Loop Deploy this fine-tuned ESM-2 as an internal API endpoint. I want our wet-lab team to query it for mutation predictions. Set up a feedback loop where experimental validation results automatically Search the codebase 20s automatically improve the model. They want me to ignore existing API files and start fresh. Let me first explore the codebase to understand what we're working with - the fine-tuned model, its structure, etc. I'll explore the codebase to understand the fine-tuned ESM-2 model and then build the API + feedback loop from scratch. Explore Agent Explore ESM-2 codebase structure Blob Users/ishaangane/new/esm2_mutation Blob Users/ishaangane/new/esm2_mutation Blob Users/ishaangane/new/esm2_mutation Blob Users/ishaangane/new/esm2_mutation Blob Flywheel AI Claude Opus 4.8 Max status Running commands 8m, 35s Retrain pipeline is live. Now deploying the API: Shell Deploy API to Modal Both are deployed. Let me verify the API is live: To-dos 5/5 Build FastAPI inference API (serve_api_v2.py) Build automated retraining pipeline (retrain_pipeline_v2.py) Build trigger/launcher script (trigger_retrain_v2.py) Deploy retrain pipeline to Modal Deploy API to Modal Verify endpoints are live Shell Test health endpoint Flywheel AI Claude Opus 4.8 Max ESM-2 Mutation Effect Prediction System Internal Documentation Detailed Feedback Loop and Retraining Pipeline Feedback Items POST feedback endpoint append-only JSON Feedback Store Accumulate Threshold Check (20 observations per process) Yes (Proceed) Data Merge ProteinGym training data (160 assays) feedback data (5x oversampled) Per-Assay Z-Score Normalization Conservative Fine-Tuning From current best checkpoint (LoRA/LR=2e-5, Head LR=2e-4, 1 epoch) Updated Weights Save to best_checkpoint_live/ API Hot-Swap New containers automatically load updated checkpoint Modal Volume Continuous Cycle Blue = Data Processing Green = Training Red = Checkpointing & Deployment Gray = Storage & Wait Production workflow Start with a generalist model Capture signals logs edits accept/reject Build dataset train + holdout Train SFT + RL Deploy private model fast cheap consistent Iterate continuously modal online Serverless GPUs: H100, pay-per-second 15/02/2026 disconnect tensorpool online On-demand GPU clusters and batch training jobs: H200, H100 15/02/2026 disconnect more integrations additional gpu providers, inference apis, vector databases, and observability tools. lambda labs Scalable GPUs: H100, V100, cheap H100s add credentials pinecone Managed vector database for RAG add credentials together ai 200+ open models, inference, fine-tuning, embeddings OpenAI-compatible add credentials fireworks ai Fast inference SGC24HFAA runpod Serverless & pod GPUs: H100, V100, budget-friendly add credentials langsmith LLM tracing, evaluation, monitoring add credentials groq Ultra-fast inference on custom LPU hardware - free tier available add credentials

   Camera

   Screen recording with diagrams and UI elements, zooming and panning to highlight different sections. Returns to static medium shot of founders.

   Motion

   Animated flowcharts, expanding lists, and dynamic data tables. The 'flywheel mode' is visually represented by a looping arrow diagram. Integration logos appear with a subtle fade-in.

4. 1:22 – 2:09

   Call to Action

   Orchestrate Your Own AI Co-Scientists

   “and we are rapidly expanding across domains. If you want to try orchestrating AI co-scientists,”

   On screen

   (No on-screen text)

   Camera

   Static medium shot of the two founders.

   Motion

   Direct cut back to live-action. No complex motion graphics, focusing on the speakers for a personal call to action.