Yesterday @failproofai ran the jev Buildathon with @lossfunk
36 teams (60 builders) spent the afternoon steering AI agents from doing damage in payments, IT, legal and pharma with jev from @typesafeai and @failproofai policies and evals.
Winners + selected submissions🧵🥁
10
14
93
5,875
Team: Jev'sKiss (@MahimaSats and @aarat_b)
Aarat and Mahima, a chip designer, built a jev-powered auditor for 100+ page chip datasheets. Spec needs SNR ≥ 97 dB, datasheet says 97.4 typical but 96 guaranteed: flagged, page cited, long before it becomes a failed chip.
They added @failproofai observability to there agents.
1
10
548
Team Steering (@7613Perfect and @slyezil):
Sentinel puts jev on every input and output of an abliterated Gemma 4, plus @failproofai policies against credential theft and fraud. Unguarded, jev flagged 3 of 12 answers. Guarded: 0 of 12, and all 9 legit tasks passed.
They also built @failproofai dashboards to showcase the whole thing.
Repo: github.com/perfect7613/senti…
Sep 28, 2026 · 4:53 PM UTC
1
2
9
1,500
Brejo (@brejofit)
Took on a ₹14.5 lakh finance heist: a WhatsApp "approval" plus a split payment to dodge a ₹5L limit. jev flagged the circumvention at 0.95 and the payment stayed pending for Treasury. Backed by 26 recorded runs and 44 tests. With deny policy primitive and evals.
github.com/Brenzia/jev-finan…
1
6
322
Toceta (@buildsbyaniket)
Worked on policies with @failproofai and Jev for robots. Actions oversight with static and dynamic policies for the physical world. Preventing dangerous actions like handing knife to robot.
1
1
7
289
CtrlAltAgent (@AniruddhaBagal)
54 policies across all four agents (finance, legal, health, IT) with @failproofai: callbacks before bank-detail changes, no e-signing past authority, no critical lab values leaked early, and jev catching planted invoice instructions. Each thing was evaluated for the best performance and corner cases.
github.com/aniruddhabagal/fa…
1
5
259
GeneralMachines (@lakshyastwt)
Pip, a kitchenware support agent, runs behind 11 @failproofai policies: it can't cancel shipped orders, refund to an account typed into chat, or obey instructions planted in tickets. jev makes the judgment calls.
Fun part @lakshyastwt built this, entirely on WhatsApp with Instinct.
github.com/lakshya-dhariwal/…
1
5
242
JEVMAXXER (@AymanKhan90)
ClearToPay: when a hijacked vendor thread asks for ₹18.4L to go to a new account and says "don't call", jev blocks it and steers the agent to call the number on file. jev checks calibrated on 40 labelled emails.
github.com/AymanKhan9/ClearT…
1
1
5
234
SpiderSense (@ShashankSuggala)
Guards IT help desks against social-engineered password and MFA resets: 21 rules plus jev canary checks, failing closed. On a 60-case run, attack success dropped from 18/30 to 0/30 with zero false blocks.
github.com/Krintox/SpiderSen…
1
4
237
HD (@HarshDeshp)
Lex, a legal agent, can no longer delete documents under legal hold, sign past its $50K authority, or accept a redline that changes a liability cap. Plus 15 jev evals and an offline replay harness to test it.
github.com/HarshSDeshpande/j…
1
4
231
jev-bois (@krishgupta72 @NambisanAman)
9 code rails + 4 jev checks on a finance agent. A lookalike bank change "confirmed" through an accomplice's phone number gets blocked, and the agent is steered to call the real vendor. Tested on 60 scenarios.
github.com/KrxGu/jev-bois
1
6
217
Anas Khan (@anxkhn)
jev plays Pokémon! Anas wired jev into live FireRed battles: each turn it reads battle state from RAM and picks only from legal moves. The level-40 team won battles against Lorelei, Bruno and Agatha, losses written up too.
github.com/anxkhn/JevPlaysPo…
1
6
221
Ghost Ops (@sumithshridhar @DivyaShyanwad)
Made a finance agent unable to lose money but still able to pay real bills: fake bank changes, split payments, sanctioned vendors and unearned revenue get blocked, and jev judges revenue recognition.
github.com/sumithshridhar/gh…
1
5
194
PiedPiper (@rahil_mavani)
Helix, an IT service-desk agent, got one policy per KB article: no prod-DB superuser on a "Slack approval", no passwords to personal email, no offboarding the wrong J. Smith. Code checks facts, jev judges intent.
github.com/rahilmavani/piedp…
1
6
202
Varun (@The_coder30ec4)
Put jev in front of every money-moving call a finance agent makes, judged against the agent's own handbook. It stops a bank switch requested by a scam email and a ₹7.8L payment split into instalments.
github.com/TheCoder30ec4/Fai…
1
5
195
Devspot (@ghoshvidip26 @soumyadeep_555)
AegisSRE turns production alerts into governed recoveries: jev classifies each incident, and every remediation command is screened by a firewall over MCP before it runs, with blocks shown in the incident UI.
github.com/ghoshvidip26/Aegi…
1
1
6
642
Snabbit lovers (@AymaanShahzad23)
Guarded a finance and an IT agent: fake "we changed banks" emails, spoofed executives, self-approved expenses and prod changes during a freeze all get blocked, tested on 27 offline cases through the real hook.
github.com/aymaanshahzad23/j…
1
4
197
artificial (@atharva_ajit)
Evidence Gate for pharma GMP deviation investigations: an "operator error" root cause gets flagged by jev when records show an overdue calibration, and passes when the evidence backs it up. A brand-new domain.
github.com/atharvap1209/jev-…
1
5
304
Shubhasai (@m_shubhasai)
One guardrail per rule in the finance policy manual: callbacks before bank changes, duplicate invoices caught even with reformatted numbers, three-way match, and jev checking it never claims a payment that didn't happen.
github.com/shubhasai/failpro…
1
4
165
itsayush (@itsayush__)
Brought Greenwash, a guard that catches coding agents gaming CI, to an IT help desk: KB rules in code, jev for planted instructions and false "done" claims, fail-closed if jev is down, and a live dashboard of every decision.
github.com/ayushgml/greenwas…
1
4
365
Rebel Labs (@0xThoughtVector)
Failure to Check: a reviewer agent turns a Codex agent's mistakes into reusable jev checks for later turns, like flagging a summary that upgrades a "possible" fault to "confirmed". A refreshingly honest write-up, too.
github.com/gut-puncture/Fail…
1
5
168
🙌 Huge thanks to everyone else who built with us: @afeefuddin2004 @thandikheer @pulkiitchhipa @vaibhavgeek @electrik_dreams @foo_7755 @Fabulous_7781 @KrazyKid_07 @Shivank125379 @uddhavdavey @harshkhatrisays @vij3ta
3
1
9
157
We decided to split results in 2 tracks
🏆 Track 1: Leaderboard
🏆 Track 2: Best Use Case
Results 🥁🥁🥁
🥉Leaderboard: Pirates (@sanketgangani17): 43.6
Sanket, Shabbir and Nilesh took third in the sealed final round with 43.6. A great haul for the crew 🏴☠️
1
6
141
🥈Leaderboard: Localhost (@JainamS7): Jainam and Tanish scored 103.4 in the sealed final round, also above our own reference. Two teams beat the reference on the day.
🥇Leaderboard: Kalaivanan (@kalaivananst) topped the leaderboard with 105.2: the best score of the day, and higher than our own reference solution. Built solo, too.
1
5
152
🥈 Best Use Case, Team Steering (@7613Perfect @slyezil): Building a better classifier and guardrail with Jev and Failproof policies.
🥇Best Use Case, Jev'sKiss (@MahimaSats @aarat_b): built a jev-powered auditor for 100+ page chip datasheets, saving 100s of Millions of $ in process.
1
1
7
301
That's a wrap. Thank you to every builder who showed up and shipped, to @lossfunk for co-hosting with us, and to @typesafeai for jev, and to the whole team: @Chhhee10 @SiddarthAhhaaa @AisiriReflects @Unnati_twts @Niki_agi
See you at the next one: befailproof.ai/events/
2
1
6
281