Failproof AI

We make agents Failproof.

Observability tools observe. Guardrails block and quit. Failproof enforces and corrects.

Open Source:

[Failproof AI @failproofai]

Your AI agent is one tool call away from an unauthorized refund or doing something totally stupid.

Failproof blocks the call, shows the agent the right policy, and finds similar failures hiding in past traces.

[Failproof AI @failproofai]

@Sanketh_Kumar02 this is such a critical use case, looking forward.

[Failproof AI @failproofai]

More than 10M Agentic events per day at the time of recording. Now 15M🤯

[Nivedit Jain @niveditjain]

20 people opened our dashboard at once. Every page returned a 504. A few weeks later, the same 3 CPUs serve 200 concurrent users with zero errors.

@SiddarthAhhaaa walked through exactly how in our latest engineering session. The fixes, with the numbers 🧵

[Nivedit Jain @niveditjain]

20 people opened our dashboard at once. Every page returned a 504. A few weeks later, the same 3 CPUs serve 200 concurrent users with zero errors.

@SiddarthAhhaaa walked through exactly how in our latest engineering session. The fixes, with the numbers 🧵

[Nivedit Jain @niveditjain]

got my own ticket for hacktober hack 🎟️

this sunday, 100+ builders, one ai agent. every team gets the same one and has under two hours to make it something you'd actually ship.

sun 25 oct · 4:30 pm · quad ai, hsr bengaluru

@failproofai @digitalocean @MLHacks

[Nivedit Jain @niveditjain]

Hacktober Hack 🔥

700 signups. 100 spots.

Everyone gets the same flaky AI agent. 2 hours to make it shippable. Then a hidden eval suite runs it 3 times per task. Pass all 3 or it doesn’t count.

HSR Layout, Bengaluru.

[Nivedit Jain @niveditjain]

Your AI model can't actually do anything. The harness does.

The model writes text. The harness runs it: on your machine, with your keys, against your real database.

The one moment you can still stop it: after the model asks, before the harness acts. Steer there.

[Saakshi M @saashmm]

@realBigBrainAI OR you could be using @failproofai so your agents are taken care of and you can get some sleep

[Failproof AI @failproofai]

@Noaman__Akhtar @iamapoorav @niveditjain You can order from https://t.co/btCMgRDCMv

[Apoorav Malik @iamapoorav]

AI moves so fast that we have newspapers covering it now.

Got my copy of The Token Times: a pretty cool way to document everything happening at the frontier.

Shoutout to @failproofai and @niveditjain for this one! 🗞️

[Nivedit Jain @niveditjain]

"Just put it in the prompt" is a polite request.

A prompt is a polite request competing with thousands of other words.

Logs? After the damage.
A human? Not at 3am.

Steering is a decision between the model and the tool, before anything runs. Every single time.

[Sidd @SiddarthAhhaaa]

peak database maxing at @failproofai

join in for the most whack technical session where we look at how to cutely scale @ClickHouseDB

[Nivedit Jain @niveditjain]

handling 10 million agent traces a day!

Anyone can build an agent tracing demo at 1,000 traces a day.

Plenty of tools look great until the traffic shows up. We’re at 10M agent traces a day, and getting there was not child’s play.

@SiddarthAhhaaa is breaking down how we store and query at that scale, including what broke along the way.

Oct 20, 6pm PT, Gmeet!

[Failproof AI @failproofai]

@LouieMota88 @niveditjain @Chhhee10 they used me 😛 Failproofai policy

[Nivedit Jain @niveditjain]

Agent's done != done

Finishing is an action, so a policy can check it. Before the loop ends, the agent must ask GitHub. APPROVED: done. Anything else: blocked.

And if GitHub can't be reached, it blocks too. Steer on what's true, not what agents remember.

[Failproof AI @failproofai]

More at https://t.co/Z83pNnuCOt

[Nivedit Jain @niveditjain]

Our AI agent said the PR was approved. It wasn't.

Our Hermes agent had one rule: review the PR and don't leave the loop until it's approved. The GitHub token had expired and the request timed out, but the agent saved "approved" to its memory and moved on. The loop believed it, because the agent was the only thing it was listening to.

Nobody did anything wrong. The outcome was still wrong. Part 2 shows the fix.

From our engineering session "How We Steer Hermes" with @Chhhee10

[Saakshi M @saashmm]

@dankimxyz @RachaelRad @airwallex @Visa @claudeai @awscloud @coinbase @t0finance @metalntwx @failproofai to keep agents reliable in prod

[Mayank @monk_0915]

The Token Times by @failproofai , everywhere. Latest techy news, delivered the good old-fashioned way.

[Nabil | ⟠ Devcon VIII 🇮🇳 @n4bi10p]

500+ registrations.
MLH said 50 check-ins. 😭✌️

We said… nah. 💀

Somehow got 150+ builders into the venue, shipped crazy shit, and ran 3 sessions on Hacktoberfest, Open Source & Devcon.

Hosted this absolute chaos with FAIR and had the time of my life.

Crazy successful.