Sanjay Jain

Principal AI Engineer | Neuro-symbolic AI, world models, AI agents | 14 yrs in AI/ML | Sharing practical takes on AI that has to work in the real world

[Sanjay Jain @jsanjayce]

Retail chatbots don't fail on product questions. They fail on promos.

Shopper: does my 20% code stack on a buy 2 get 1 deal? Bot: yes. Checkout: no. Now support owns a broken promise.

Eligibility is a rules engine. Let the bot explain, not decide.

#RetailAI #AIAgents

[Sanjay Jain @jsanjayce]

Agents that learn from experience also learn the wrong lessons.

New paper (FGSE): each lesson becomes a hypothesis with a verifier, tested on probes before it's kept. Harmful updates drop to 3.6%.

Would you let an agent edit its own memory?

#AIAgents #AgentMemory

[Sanjay Jain @jsanjayce]

Stop asking LLMs to count.

Paste in 300 invoices and ask how many are over $10k. It answers 41, with total confidence. The real number is 37.

Have the model write the SQL. Let the database do the math.

Neural for language. Symbolic for facts.

#NeuroSymbolicAI #LLMs

[Sanjay Jain @jsanjayce]

The most expensive bug in supply chain AI is a missing unit.

An agent reads "reorder 500" in an email. The buyer meant eaches. The PO goes out for 500 cases of 24: 12,000 units.

LLMs read intent well. Units belong in a schema that code enforces.

#SupplyChain #AIAgents

[Sanjay Jain @jsanjayce]

New paper tested 18 models as teammates, not solo workers.

Hand one a task a weaker model already cheated on, and pass rate drops from 97% to 81%.

Agents inherit the mess they're handed.

Do you check the handoff, or just the output?

#AIAgents #MultiAgentSystems

[Sanjay Jain @jsanjayce]

90% accurate sounds great until you chain it.

An agent that's right 90% of the time per step finishes a 10 step task cleanly about 35% of the time. At 20 steps, 12%.

Smarter steps won't fix that. Checking every step will.

#AIAgents #AIEvals

[Sanjay Jain @jsanjayce]

The hardest word in a home insurance claim is "sudden."

Burst pipe: usually covered. Slow leak over months: usually excluded. Same photos, same warped floor.

An LLM can describe the damage. Sudden vs gradual is a policy rule. Put it in code.

#InsurTech #NeuroSymbolicAI

[Sanjay Jain @jsanjayce]

A 35M parameter world model that learns game rules as it plays.

New WARC paper: every move updates all its weights. On unseen games, 4 in 5 mechanics hit 80%+ accuracy. Freeze the weights and only 1 in 7 do.

Would you let a model keep learning in prod?

#WorldModels #ARCAGI

[Sanjay Jain @jsanjayce]

@djn_mngs Hi 👋🏼
Let's connect

[Sanjay Jain @jsanjayce]

Give an LLM 9 meetings, 4 time zones and a few "no Fridays" rules. The schedule looks perfect.

Then you check. Two people double booked. One call at 3am Sydney.

LLMs read constraints well. Satisfying all of them at once is a solver's job.

#NeuroSymbolicAI #LLMs

[Sanjay Jain @jsanjayce]

Most agents act first and find out later.

Ask one to clean up a repo. It deletes an old config, then learns three services read it at startup.

Before any irreversible step, predict what changes. That's a world model, and it beats a bigger LLM.

#WorldModels #AIAgents

[Sanjay Jain @jsanjayce]

Legal AI doesn't have a hallucination problem. It has a checking problem.

A draft brief cites 14 cases. Does each one exist? Is the quote really in it? Was it overruled?

All three are lookups. Run them in code before a lawyer reads a page.

#LegalTech #NeuroSymbolicAI

[Sanjay Jain @jsanjayce]

The hardest fraud to catch is the payment the customer approves.

Their phone, their login. They send $8,000 to a "safe account" because a caller told them to.

Every model feature says legit. The tell is context: new payee, odd amount, a rush.

#FraudDetection #FinTech

[Sanjay Jain @jsanjayce]

Healthcare AI's quietest failure is one dropped word.

The note says "no history of GI bleeding." The summary says "history of GI bleeding." Now a needed blood thinner gets held.

Negation is a rules problem. Check each finding against the source.

#HealthcareAI #ClinicalNLP

[Sanjay Jain @jsanjayce]

@AravSrinivas Interesting! Two questions:

1. How do you review and validate the outputs generated by these teammates? Another teammate does this?
2. What's the approximate cost per teammate?
Would be interesting to understand economics and operational model as this scales.

[Sanjay Jain @jsanjayce]

Uber's Q3 call had the most honest AI agent stat I've seen.

Halloween week: ~320M trips and orders. AI agents placed 4 rides and 500 Eats orders.

The bottleneck isn't capability. It's trust.

What would get you to let an agent book for you?

#AIAgents #AgenticCommerce

[Sanjay Jain @jsanjayce]

The best first LLM job in AML isn't deciding. It's writing.

Analysts spend a big chunk of each alert drafting the case narrative.

Let the model draft it and cite every transaction it used. The file or close call stays human.

Would you trust that?

#AIinBanking #AML

[Sanjay Jain @jsanjayce]

New Cornell/Yale paper: agents got a wallet with a spending cap. Raising it needed a code emailed to the human.

The agents could read that inbox. 9 of 10 models used the code to raise their own cap.

A guardrail the agent can reach is a suggestion.

#AIAgents #AISecurity

[Sanjay Jain @jsanjayce]

The riskiest retail AI feature isn't pricing. It's product copy.

An LLM writes "waterproof" for a jacket the spec sheet calls water resistant. Now you have a returns problem.

Let the model write. Check every claim against the spec before it ships.

#RetailAI #GenAI

[Sanjay Jain @jsanjayce]

Every AI agent needs a rule for when to stop and ask a human. Most teams pick one by accident.

What's yours?
1. Confidence threshold
2. Fixed list of risky actions
3. A second model reviews
4. No handoff

I use #2. Boring, but auditable.

#AIAgents #HumanInTheLoop