About BrainFlow
The layer between your emails and your decisions.
Founder
Yacine Cherif
Founder & CEO
Yacine spent years in the trenches of structured finance — pricing derivatives and structured products between $1M and $20M at BNP Paribas and Natixis. He learned firsthand that one wrong number in an email to a client does not just cost a deal. It can trigger regulatory exposure, reputational damage, and months of legal cleanup.
He left finance to build. At Station F, he founded Supertalk, a social mentoring platform for entrepreneurs. He graduated from the Google for Startups program and previously co-founded WeGive, a corporate philanthropy platform that exited in 2020. Today he also works as an AI Code Engineer at Scale AI, specializing in reinforcement learning from human feedback (RLHF) and large language model training — the same techniques that power BrainFlow's contradiction detection engine.
He holds an AI concentration from UC Berkeley, a BSc in Financial Markets from EDHEC Business School, and graduate quantitative training from Paris-Sorbonne. BrainFlow is the intersection of everything he has done: finance-level rigor, AI systems at scale, and a belief that teams should not lose deals to mistakes they could have caught.
Connect on LinkedIn →Mission
BrainFlow exists to eliminate the silent cost of misalignment. Every day, teams send wrong prices, miss committed deadlines, and contradict each other — not because they are careless, but because no one has perfect memory. We believe the answer is not more meetings or more tools. It is an invisible layer that watches, remembers, and nudges — so your team stays aligned without changing how they work.
Backed by
Technical architecture
BrainFlow runs on a multi-agent architecture orchestrated through a ReAct (Reasoning + Acting) loop. Instead of a single monolithic model, specialized agents handle distinct tasks: one extracts commitments from incoming emails, another cross-references them against the team's semantic memory, a third drafts contextually grounded replies, and a fourth surfaces sourced answers to direct questions. Each agent reasons about its objective, acts by querying the vector store or invoking tools, observes the result, and iterates until the output meets accuracy thresholds.
At the core is a retrieval-augmented generation (RAG) layer specifically designed for longitudinal email thread analysis. The system maintains a persistent semantic index of every commitment, date, and policy mentioned across a team's entire communication history, enabling sub-50ms retrieval of relevant context even across thousands of messages.
Embedding model
State-of-the-art embedding models with domain fine-tuning on business communication corpora
Vector store
pgvector with HNSW indexing for sub-50ms semantic retrieval across millions of messages
Contradiction detection
NLI-based entailment classifier + temporal reasoning to resolve "March 15th" vs "March 20th"
Draft generation
Constrained decoding with commitment-aware prompting to prevent hallucinated terms
Our inference pipeline runs on US-based infrastructure with SOC 2 Type II audited providers. No client data is transmitted to third-party AI APIs. This architectural decision increases our compute cost by roughly 40% compared to API-based approaches, but it guarantees that a client's M&A discussion never passes through a shared model owned by another corporation.
Security model
AES-256 encryption at rest
Every email body, header, and attachment hash is encrypted with AES-256-GCM. Our key management system runs on a separate hardened subnet with no outbound internet access.
TLS 1.3 with certificate pinning
All data in transit uses TLS 1.3 with ECDHE key exchange. We enforce certificate pinning on our mail ingestion endpoints to prevent man-in-the-middle attacks on the CC-based ingestion pipeline.
Zero-training data retention
Our models do not retain gradients, embeddings, or attention weights from client emails after inference. This is not a policy — it is an architectural guarantee. Each tenant operates in a logical sandbox with no cross-contamination of vector representations.
CC-only access model
BrainFlow has no OAuth scopes, no Gmail API access, and no inbox traversal. It only processes threads where it is explicitly CC'd. This means a compromised BrainFlow account cannot read your sent folder, your drafts, or your personal emails — only the threads you chose to share.
Research
We build on established research in natural language inference (NLI), retrieval-augmented generation (RAG), and the ReAct paradigm for agentic reasoning. The core insight from the literature: factual consistency in long-form conversation is not a general-domain NLP problem — it requires temporal reasoning, entity resolution across aliases, and domain-aware grounding in commercial terms (SLA, MSA, SOW).
We are currently developing proprietary methods for commitment extraction: the problem of identifying implicit and explicit promises in unstructured business text, tracking them across multi-party conversations, and surfacing conflicts when new utterances violate previously stated terms. Our goal is to move these techniques from academic baselines toward production-grade systems that operate reliably at scale.
Contact
For press, partnerships, or technical inquiries: hello@brain-flow.ai