Ashish Vishwakarma

Senior AI Researcher | Building AI that can prove it worked 🛡️ | Multi-Agent Systems · RAG · LLMs | Creator of Veridian (PyPI) | IIT Bhilai | @SigmaRedAI

[Ashish Vishwakarma @ashishcse31]

Good call.

The “repair until green” loop is the most relatable failure mode: every engineer understands an agent repeatedly editing, rerunning tests, and drifting without a real stop condition.

Plan is to turn it into:
bad loop -> Loop Doctor -> bounded repair contract -> passing evidence.

[Ashish Vishwakarma @ashishcse31]

Shipped a major update to LoopRight.

It’s a portable Agent Skill for Codex, Claude Code, and coding agents that helps design/review/repair loops that are bounded, observable, failure-aware, and backed by evidence.

Loops are where production risk hides:

“just retry”
“keep polling”
“fan out everything”
“iterate until it works”

LoopRight now includes:

- Loop Doctor
- CLI: scan/doctor/validate/catalog
- SARIF code-scanning output
- pre-commit hook
- pattern catalog
- runnable proof examples
- scanner benchmarks
- realistic field-guide examples

Goal: make loops explicit before they become incidents.

Repo:
https://t.co/nMAkYHOoL0

#AIEngineering #DevTools #CodingAgents

[Ashish Vishwakarma @ashishcse31]

Introducing LoopRight : an open-source Agent Skill for Codex, Claude Code, and other compatible coding agents.

Before an agent writes a retry loop, polls an API, tunes a model, runs a benchmark, or keeps repairing code, it should answer:

• What proves progress?
• What is the budget?
• When should it stop?
• Which failures are safe to retry?
• What evidence proves completion?
Complex example: ML model tuning
Instead of prompting:
“Keep tuning the model until accuracy improves.”
Use LoopRight:

$loopright Review and implement the tuning loop.

Objective:
Maximize validation F1.

Constraints:
- Use existing Optuna integration
- Maximum 50 trials
- Maximum runtime: 2 hours
- Do not tune against the test set
- Maximum train-validation gap: 0.05
- Enable pruning and early stopping
- Preserve seeds, dataset version, and code version

Completion evidence:
- Baseline vs best-trial comparison
- Holdout test evaluated only after tuning
- Best parameters saved
- Best model artifact saved
- Relevant tests pass

The goal is not just to make agents iterate more.

It is to make them iterate with boundaries, measurable progress, safe failure handling, and proof of completion.

Fewer runaway retries.
Less wasted compute.
Safer side effects.
No false “done.”

https://t.co/nMAkYHOoL0
#AgentSkills #Codex #ClaudeCode #OpenSource #MLOps #SoftwareEngineering

[Ashish Vishwakarma @ashishcse31]

Built and pushed a release-facing cleanup to Veridian main:

-> clearer README positioning (deterministic
verification + replay-safe runtime)
-> aligned PyPI metadata/keywords
-> cleaned package-level docs links

Goal: reduce trust friction and make adoption clearer for production teams.

Repo: https://t.co/mwtShWk0oi
#AIEngineering #AgentReliability #OpenSource #Python

[Ashish Vishwakarma @ashishcse31]

We just shipped a new Veridian example: mission-critical wire fraud release control.

What it proves (with tests):

High-risk wires pause for dual approval
sanctions-hit wires are blocked. Release side effects are replay-safe (no duplicate transfer calls)
This is a production-pattern reference implementation, not a live banking rails integration yet.

pytest examples/12_wire_fraud_release_review/test_pipeline.py -q
python examples/12_wire_fraud_release_review/pipeline.py
#AI #LLMOps #Fintech #Reliability #OpenSource41

https://t.co/mwtShWk0oi

[Ashish Vishwakarma @ashishcse31]

Veridian updates:🚀

The verification layer for AI agents — correctness enforced by the runtime, not by prompting.

What's new:
→ Pause/resume that survives crashes
→ Zero-duplicate LLM calls via activity journal
→ Graph execution engine with verified edges
→ LangGraph + CrewAI adapters (drop-in reliability)
→ Multi-tenant isolation (budget, rate limit, data)
→ Operator plane: approval queue, DLQ triage, replay CLI
→ Plugin SDK with certification matrix
→ SLO engine + PII trace filtering

Python 3.11+. MIT.
https://t.co/2culzJgqia
#AI #OpenSource #AIAgents #Python #LLM #DevTools

[Ashish Vishwakarma @ashishcse31]

Just shipped Veridian v0.2.0 🚀

Verification-as-infrastructure for AI agents. Not another wrapper — correctness enforced by the runtime.

What's new:
→ Resumable pause/resume (survives crashes)
→ Zero-duplicate LLM calls via activity journal
→ LangGraph + CrewAI adapters with verified edges
→ Multi-tenant isolation (budget, rate limit, data)
→ Replay/diff CLI for operator debugging
→ API surface cut from 123 → 40 stable symbols
1,625 tests. MIT. Python 3.11+.
The bet: frameworks that win in 2026-27 make agent reliability an infrastructure guarantee, not a prompting prayer.
https://t.co/2culzJgqia
#AI #OpenSource #AIAgents #Python #LLM

[Ashish Vishwakarma @ashishcse31]

@MilksandMatcha Building a deterministic verification infrastructure for autonomous AI agents. https://t.co/aEuWr6jPz9

[Ashish Vishwakarma @ashishcse31]

I just published AI Agents Don’t Need Better Models. They Need Boring Infrastructure.

https://t.co/ZhLQKDe8Vd

https://t.co/aEuWr6jPz9

[Ashish Vishwakarma @ashishcse31]

March 31, 2026.

Anthropic leaked 512,000 lines of source code via a forgotten .map file in npm.

Mercor AI lost 4TB to Lapsus$ — alleged vector: a dev pasted production
credentials into Claude.

The AI industry's security problem is now eating itself.

Both failures had the same root cause:
agents running in production with zero verification infrastructure.

No artifact fingerprinting. No secret sanitization. No proof chain.
No external verifier. Just "the model said it's fine."

Every database has ACID guarantees.
Every CI pipeline has automated tests.
Agent frameworks ship with none of this — and we accept it.

Check your lockfile for [email protected] right now.
Then ask yourself what guarantees your agent pipeline actually has.

https://t.co/2culzJgqia

#AgentSafety #AIInfrastructure #Anthropic #Mercor #InfoSec

[Ashish Vishwakarma @ashishcse31]

Veridian Phase 6b shipped — anti-misevolution safety layer 🛡️

3 new verifiers protecting against AI agent self-corruption:

→ ToolSafetyVerifier — AST-based static analysis blocks eval/exec, shell injection, blocked imports in agent-generated code
→ MemoryIntegrityVerifier — detects reward hacking, prompt injection, contradiction & numeric drift in memory updates
→ VerifierIntegrityChecker — SHA-256 fingerprints the verification chain at run start, catches tampering at run end

https://t.co/2culzJgqia

[Ashish Vishwakarma @ashishcse31]

Veridian now has a CLI.

veridian init — create a new ledger
veridian run — execute tasks through verification veridian status — Rich table with pass/fail/cost stats
and more ....

Dry-run mode built in.
586 tests.
Strictly typed.

MIT licensed.

https://t.co/PrDIAg4ZwG

[Ashish Vishwakarma @ashishcse31]

New in Veridian: DriftDetectorHook

Tracks per-run metrics — verification pass/fail rates per verifier, confidence distributions, retry rates, and token consumption. Persists snapshots as JSONL. Compares the current run against a configurable historical window.

Detection uses Bayesian Beta lower-bound (same formula as our SkillLibrary scoring) + z-score for continuous metrics. Flags when the magnitude exceeds the threshold, and statistical significance holds.

Catches: pass rate drops, confidence degradation, retry spikes, token burn increase, failure mode clustering.

One line to enable:
drift_history_file = "drift.jsonl"

Read-only hook. Atomic persistence. 24 new tests. Zero external deps.

pip install veridian-ai

https://t.co/2culzJgqia

[Ashish Vishwakarma @ashishcse31]

Every AI agent framework gives you a loop. None gives you a guarantee.

So I built Veridian — deterministic verification for autonomous agents. Tasks can't be marked DONE unless they pass a Python verifier. The LLM can't override this.

Public beta. MIT licensed.
🔗 https://t.co/2culzJgqia

[Ashish Vishwakarma @ashishcse31]

@Suryanshti777 Book

[Ashish Vishwakarma @ashishcse31]

@SiddharthAulia Om Shanti.

[Ashish Vishwakarma @ashishcse31]

@airtelindia
Mobile Internet speed very slow(50kbps apx). Reported same to customer support, problem persist the same. Please look into it. PINCODE:222139

[IIT Bhilai @IIT_Bhilai]

Here are a few tips to keep yourself engaged and mentally happy during the quarantine
@IIT_Bhilai
@HRDMinistry
@MoHFW_INDIA
@PIBHRD

[IIT Bhilai @IIT_Bhilai]

Saksham Bhushan & Debajyoti Halder, students of @IIT_Bhilai created the Website: https://t.co/8xasnTVvXf
which helps to visualize the current status of coronavirus affected states in India. Please share and stay safe.

@HRDMinistry
@MoHFW_INDIA

[IIT Bhilai @IIT_Bhilai]

In response to the COVID-19 outbreak, an awareness session was conducted by the Health center at @IIT_Bhilai to educate the housekeeping staff and security with basic hygiene standards.
@HRDMinistry
@PIBHRD
@V2019N