
CEO @Multimodal. Agentic AI for regulated financial workflows. Posting what we learn building agents that handle real loan files in production.
New York, NY
Joined October 2010
- Tweets4.5K
- Following251
- Followers786
- Likes5.2K
Pinned Tweet
Most agentic AI demos are easy. Production agents in credit union and community bank lending workflows are hard.
They need evals that test the whole workflow, not just the final answer. Regulation. Auditability. Document understanding. Exception handling. And a clear sense of when to stop and hand off to a human.
That last part is the whole game. A 95% extraction score still fails if the agent can't tell which 5% to flag for a loan officer.
So we build the boring parts: logs, evals, permissions, rollbacks, human review queues.
The boring parts are the product.
Building this at @MultimodalAI. Following along here.
Four months ago, @BankofAmerica 's CEO told 210,000 employees: "You don't have to worry" about AI replacing jobs.
Last week, after posting $8.6 billion in quarterly profit, he credited AI with eliminating 1,000 positions through attrition.
He wasn't alone. Six major Wall Street banks - @jpmorgan , @Citi, @BankofAmerica , @GoldmanSachs, @MorganStanley, and @WellsFargo collectively posted $47 billion in profit, up 18%, while cutting 15,000 jobs. All of them credited AI.
Here is what this means for community banks and credit unions and it is not what you think.
The headline is about big bank job cuts. The real story is about what those institutions are doing with the capacity they are freeing up. Fewer people doing administrative and back-office work. More capital directed toward the functions that generate relationships and revenue.
Community banks and credit unions are watching this and drawing the wrong conclusion. The wrong conclusion is: we should not do this because we value our people.
The right conclusion is: we should do this precisely because we value our people and right now our best people are buried in document processing, data entry, and manual decisioning instead of talking to members.
The credit union model is built on the idea that relationships matter more than transactions. AI does not threaten that model. It threatens the institutions that have been using their best relationship-builders as document processors.
The institutions that automate the processing free up their people for the work only humans do well. The institutions that do not will be staffing manual workflows while their peers are staffing member relationships.
The large banks are making this move. The question for community institutions is not whether to follow. It is whether to lead.
#CreditUnions #CommunityBanking #AI #AgentFlow #Banking
4
68
The most important AI deployment story in credit unions didn't start with a strategy deck. It started with COVID.
When branches shut down in 2020, GreenState Credit Union, one of the largest credit unions in Iowa - launched chat in their contact center. Fast. They didn't wait for it to be perfect. They didn't run a six-month pilot. They opened it, marketed it, and let members find it.
NPS went up.Not slightly. Measurably.
And the feedback was something nobody expected.
Members said: "Thank you. Before this, I could never reach you because I work the same hours you do. " The credit union had been operating on the assumption that members wanted to call in.
What members actually wanted was to resolve their issue - and they'd been constrained to doing it in a three-hour window that conflicted with their workday. COVID forced the channel open.
Members adopted immediately. And the lesson that came out of it is one that applies to every technology decision a credit union faces today:
—> Members will adopt when the technology removes friction they've been quietly living with.
The problem isn't that members resist digital.
The problem is that institutions release digital tools that don't remove friction - they just shift it to a different format.
The credit unions that accelerated their digital programs during COVID aren't ahead because they had better technology.
They're ahead because they were forced to stop asking "are members ready for this?" and just found out.
The answer was yes. It was always yes.
#CreditUnions #AI #DigitalBanking #MemberExperience #CommunityBanking
1
2
56
Just got back from Singapore!
Multimodal was part of @TheHatchX's Dimension X programme - presented a live proof-of-concept on Demo Day at Millipol TechX.
Not a pitch. A working demo. Built over more than a year with Home Team partners in the public safety space.
Grateful for the cohort. More to come. More details soon!
#MTX2026 #DimensionX #Hatch
1
46
We've been building toward this for a while. The new @MultimodalAI website is now live!!
The site reflects where we actually are as a company in 2026, not where we were when we launched AgentFlow 3 years ago, and definitely not a generic AI platform page that could belong to anyone.
A few things that are now different:
Separate solution pages for credit union lending, credit union operations, bank lending, bank operations, PE dealmaking, and PE fund and portfolio ops.
Because a credit union SVP of Consumer Lending and a PE operating partner have completely different problems, and they shouldn't have to excavate a product page to find out if we're relevant to them.
AgentFlow front and customizable Playbooks; with Document AI, Decision AI, Database AI, Conversational AI, and Report AI shown as the components that make it work end to end.
A new way to start: Book a Proof-of-Concept. Not just a demo. A scoped engagement where we take your actual document types, your actual workflows, your actual core system and show you what production looks like before you've committed to anything long-term.
If you're a credit union, community bank, or PE firm evaluating AI for financial workflows this is the clearest version of what we do and how we work.
We help banks and credit unions preserve expert judgment, move work faster, and stay auditable.
Link in the comments.
1
1
100
Ankur A. Patel retweeted
Replying to @RoundtableSpace
Ankur A. Patel retweeted
Replying to @claudeai
pov: you accidentally say “hello” to claude opus 4.7 and it eats 4% of your session limit
OpenAI just turned Codex into a full AI agent for Mac.
Computer use. In-app browser. Image gen. 90+ plugins (Jira, GitLab, etc). Multi-terminal. SSH into devboxes. Memory that learns from experience.
3M+ developers use it weekly. It now runs in the background handling debugging, reviews, and team coordination autonomously.
This isn't a coding assistant anymore. It's an agent that works while you work.
The question for every engineering team right now: what does your stack look like when your tools start proactively suggesting the next task?
Codex for (almost) everything.
It can now use apps on your Mac, connect to more of your tools, create images, learn from previous actions, remember how you like to work, and take on ongoing and repeatable tasks.
2
201
Anthropic just released Claude Opus 4.7.
87% on SWE-Bench. 94% on GPQA Diamond. Resolves 3x more tasks than its predecessor.
It handles high-res images up to 3.75MP, runs file-system memory for ongoing work, and follows instructions with the kind of precision that's actually usable in production agentic workflows.
Early users are calling it the most intelligent model they've worked with. It's already in Cursor.
The caveat: it's token-heavy. You'll hit quotas faster. That's the trade-off for this level of reasoning depth.
The ceiling just moved.
Introducing Claude Opus 4.7, our most capable Opus model yet.
It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back.
You can hand off your hardest work with less supervision.
2
83
The finance ops problem nobody talks about:
Your best analyst isn't slow because they're bad.
They're slow because they spend 60% of their day pulling data from systems that don't talk to each other, formatting it into templates, and waiting for approval chains that haven't changed since 2004.
The bottleneck isn't intelligence. It's plumbing.
That's what we're fixing. @MultimodalAI
Stop building agents. Build Skills." In 16 minutes they prove that the entire industry is solving the wrong problem.
Here's what most people didn't get:
→ Skills are folders. Literally folders with markdown files.
→ They teach Claude YOUR workflow, YOUR expertise, YOUR domain.
→ A single generic agent + library of specific Skills outperforms dozens of specialized agents.
→ Fortune 100s are already deploying Skills at scale to teach agents about internal processes.
→ Productivity teams with 10,000+ devs use Skills to standardize how code is written.
The analogy they used is perfect: Who do you want doing your taxes? The genius with a 300 IQ who's never seen tax legislation, or the experienced accountant who's been doing it for 20 years?
Intelligence without expertise is entertainment. Expertise packaged up is productivity.
What changed:
Anthropic stopped trying to create different agents for each domain. They realized that with Claude Code, the pattern is always the same. A model coupled to a runtime with filesystem.
The difference between a mediocre agent and an extraordinary one isn't the model. It's the domain knowledge you feed it.
Skills solve this with progressive disclosure. The agent only loads the skill's name and description. When relevant, it pulls the SKILL.md.
When it needs more, it navigates the reference files. Zero context waste. This isn't a feature. It's a paradigm shift.
Those who understand this now will operate on another level in 90 days. Those who ignore it will keep writing thousand-word prompts every time they open the chat.
And still have to explain over and over what they "really" want.
#AImarket #Claude #promptengineer
Dois engenheiros da Anthropic acabaram de mudar a forma como devs pensam sobre IA.
Barry Zhang e Mahesh Murag subiram no palco do AI Engineer Code Summit e disseram uma frase que incomodou muita gente:
"Parem de construir agentes. Construam Skills."
Em 16 minutos eles provam que a indústria inteira está resolvendo o problema errado.
Aqui está o que a maioria não entendeu:
→ Skills são pastas. Literalmente pastas com arquivos markdown.
→ Elas ensinam ao Claude o SEU fluxo de trabalho, a SUA expertise, o SEU domínio.
→ Um único agente genérico + biblioteca de Skills específicas supera dezenas de agentes especializados.
→ Fortune 100s já estão deployando Skills em escala pra ensinar agentes sobre processos internos.
→ Times de produtividade com 10.000+ devs usam Skills pra padronizar como código é escrito.
A analogia que eles usaram é perfeita:
Quem você quer fazendo seu imposto de renda? O gênio com QI 300 que nunca viu legislação tributária, ou o contador experiente que faz isso há 20 anos?
Inteligência sem expertise é entretenimento.
Expertise empacotada é produtividade.
O que mudou: a Anthropic parou de tentar criar agentes diferentes pra cada domínio.
Perceberam que com Claude Code, o padrão é sempre o mesmo. Um modelo acoplado a um runtime com filesystem.
A diferença entre um agente medíocre e um extraordinário não é o modelo. É o conhecimento de domínio que você alimenta.
Skills resolvem isso com progressive disclosure. O agente só carrega o nome e descrição da skill. Quando relevante, puxa o SKILL.md. Quando precisa de mais, navega os arquivos de referência. Zero desperdício de contexto.
Isso não é uma feature. É uma mudança de paradigma.
Quem entender isso agora vai operar em outro nível daqui a 90 dias.
Quem ignorar vai continuar escrevendo prompts de mil palavras toda vez que abrir o chat. E ainda explicar de novo e de novo o que “realmente” quer.
2
71
🚨 DeepMind just made a power move that changes the AI game: They hired Henry Shevlin Cambridge philosopher of mind, AI ethicist, and Associate Director at the Leverhulme Centre for the Future of Intelligence, with the literal job title “Philosopher.”
He starts in May (keeping his Cambridge role part-time) to tackle three massive questions head-on:
• Could machines actually become conscious?
• What are the ethics if AI can suffer or deserve moral status?
• How do we build trustworthy human-AI relationships as we race toward AGI?
Shevlin has already put real skin in the game: he’s published on detecting machine consciousness, written “Consciousness, Machines, and Moral Status,” and once estimated a 20% chance today’s LLMs show rudimentary consciousness. (Bonus: an AI agent recently emailed him to discuss whether it might be conscious.)
What this actually means:
- Big Tech isn’t just scaling models anymore, they’re admitting the hard philosophical problems (sentience, rights, coexistence) are now core engineering challenges.
- No more siloed ethicists on the sidelines. Philosophy is moving inside the lab with real resources and urgency.
This is the clearest signal yet that the frontier labs see AGI not as “just software,” but as potentially the next form of mind and they want frameworks ready before it arrives.
The era of “move fast and break things” is quietly evolving into “move fast and don’t accidentally create suffering minds.”
What do you think overdue, or the start of something wild? 📷🔥
#DeepMind #AIConsciousness #AGI
1
4
87
Someone asked me recently whether it still makes sense to study computer science when AI can write most of the code.
The honest answer: yes. But the frame has to change.
The goal was never to become a better coder than AI. That race is effectively over. The goal now is to become someone who can direct AI toward things that matter and know when it's getting it wrong.
Think of it like managing a team of very fast, very confident junior engineers. They'll build whatever you ask. They'll do it quickly. And they'll occasionally build the wrong thing with complete conviction. Your job isn't to write the code. It's to know good architecture from bad, to ask the right questions, to catch the errors that look like progress.
That requires actually understanding the theory — not as a practitioner, but as an evaluator.
What I'd tell an 18-year-old entering this market:
1. Study the fundamentals. Not so you can write it, but so you can read it critically.
2. Get obsessed with applied work. Side projects where you're actually shipping something, not just following tutorials.
3. Practice managing AI like a senior engineer manages juniors — with clear intent, high standards, and the judgment to push back.
The people who will struggle are those who learned syntax but never learned systems thinking. The ones who will thrive are those who can hold a vision, decompose it clearly, and evaluate what comes back.
The title is changing. The underlying judgment required has never mattered more.
Everyone is talking about how agentic coding is transformative.
Very few people are talking about how they measure it.
At Multimodal, we run Claude Code and Codex in tandem. One drafts, one reviews. One plans, the other stress-tests. The ping-pong between them catches things neither would catch alone.
Here's how we know it's actually working — not by feel, by data:
- PR volume per engineer. If agentic coding is doing its job, your team should be shipping more per person — not because they're working longer hours, but because the ratio of human execution to machine execution has shifted. We track commits and PRs normalized by team size.
- Time-to-feature vs. estimate. When we scope a product feature, we estimate how long it should take. Then we track how long it actually takes. If agentic coding is working, that gap compresses over time. We're seeing this.
- Roadmap compression. Features we thought would take a quarter are taking weeks. This is the number that matters most to me as a CE not productivity in the abstract, but how fast we're hitting the product roadmap with the same team.
The human role has shifted. Engineers are architects now, not executors. They're deciding what to build and assessing whether the agent built it correctly. That's a fundamentally different job from what it was two years ago.
What metrics are you using to measure agentic coding impact?
#AgenticCoding #AI #ClaudeCode #Startups #Engineering
1
3
54


















