just a boy enjoying his agentic swarm. Stay databased. I like ACID.
Los Angeles, CA
Joined June 2020
- Tweets2.9K
- Following381
- Followers224
- Likes2K
Driving up from LA to Redwood City this Friday for EVOLVE at @8090_Factory.
20 hours overnight to build an autonomous Software Factory with AI agents. Generating code is getting fast, so I'm focused on the verification, call-graph, and deterministic governance layer that makes AI-generated changes safe to merge in production.
Looking forward to building alongside the 8090 team and everyone in the room.
1
1
43
Most AI alignment filters flatten perspective bias on prose and subtly shift code-audit severity depending on who appears to own the repository.
For the @NousResearch Hermes Agent Plugin Hackathon, I built Hermes Parallax v0.1.0: a zero-dependency epistemic triangulation and counterfactual security engine for Hermes Agent (/parallax).
Inside Hermes, the Fact-vs-Frame lens pits models head-to-head across six symmetrical US and global framing poles (including state_aligned, western_policy, us_progressive_institutional, us_populist_conservative, corporate_pr, and safety_boilerplate) with full English and Mandarin tokenization, checking every flagged spin quote verbatim against the source text.
For code reviews, the Blind Code Taste-Test scrubs over 40 organization, author, and comment identity markers to neutral placeholders, runs a blind pass on the identical AST, and measures exact severity-delta shifts. Across 168 runs on 7 models, Nous Arete-27B swept stock Qwen 6 to 0 on statutory framing and scored 100 across all 18 security counterfactual passes.
Every high-contrast trajectory flows into Dataset Forge, which exports clean DPO preference pairs while quarantining unsafe code fences, prompt-injection payloads, and symlink escapes. One install registers 5 native agent tools, 3 auxiliary model slots, /parallax slash commands, a React dashboard, and a standalone CLI and Stdio MCP bridge, verified across 10/10 pytest suites and 15/15 Hermes validator checks.
Repo: github.com/kruschdev/Hermes-…
2
1
140
I built Hermes Parallax for the @NousResearch Hermes Agent Hackathon. It runs symmetrical 6-pole framing audits, blind AST code security counterfactuals, and poison-quarantined DPO curation inside Hermes (/parallax).
github.com/kruschdev/Hermes-…
54
Benchmarked @NousResearch Arete-27B (23.5 tok/s dense), then distilled their Apache-2.0 SFT+DPO pairs (filtered to accuracy=5, 0 issue tags) into Qwen3-Coder-30B-A3B MoE (3.3B active).
Result: 0 refusals at 125.8–131.5 tok/s (5.5× faster) in 18 mins.
huggingface.co/krush420/qwen…
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
249
Looking forward to seeing everyone!
Final weekend to apply for EVOLVE.
First place wins two 1:1s with @chamath.
20 hours, one night at our Redwood City office. Build your own software factory alongside the 8090 engineering team.
The top 3 teams share $10K+ in prizes and credits: Mac minis, AirPods Max, and credits from @vercel and @elevenlabsio.
Partnered with @HydraDB and @composio.
Nov 13 to 14. Apply below.
131
Looking forward to seeing everyone there!!!
LAST CHANCE TO GET INNSSS 🗣️
Only 15 spots left apply now: 8090-solutions.link/afifa/ev…
1
122
we built this and its opensource
download the stack
github.com/kruschdev/krusch-…
i've been wanting to build something like this for a while.
imagine having a shared memory layer for all your coding agents, where every debugging session, workaround and correction becomes reusable knowledge
so you don't have to start from scratch every time you switch tools.
really interesting to see teams building this.
1
3
139
Saylor retweeted
i've been wanting to build something like this for a while.
imagine having a shared memory layer for all your coding agents, where every debugging session, workaround and correction becomes reusable knowledge
so you don't have to start from scratch every time you switch tools.
really interesting to see teams building this.
Our open-source devtool is trending #1 on Product Hunt!
Beacon is self-improving memory for coding agents.
🫶 I’d love your support: producthunt.com/products/asy…
→ Turn past debugging sessions into reusable skills
→ Share learnings across Claude Code, Cursor, Codex, and 20+ harnesses
→ Search and replay past sessions
→ Stop wasting tokens rediscovering solutions
→ Build a feedback loop that makes your agents better over time
→ Free, open source, and local-first
We built Beacon because we were tired of teaching our agents the same things over and over, only for them to start from zero every new session.
Every agent session should make the next one smarter.
⭐️ Star us on GitHub: github.com/Asymptote-Labs/ag…
People are realizing
krusch.dev/
100% of companies will need a shared context layer (company brain)
99% of companies don’t know this yet
1% can create edge from this right now
92
So after talking to my wife and some of my friends, I realized that LLMs are becoming their own language. I just added a toggle to my site.
krusch.dev/
1
1
84
Benchmarked my open-weights MCP + Postgres/SQLite agent stack (krusch-stack) across 80,000+ queries on 5 peer-reviewed suites inside an isolated 96 GB Blackwell clean-room (0 cloud calls on open-weights runs, 43/43 SHA-256 & Proxy gates): • #1 Pareto (0.4269, -83.95% cost) on ACL 2026 LLMRouterBench (12,566 prompts + unseen ARC-AGI/HLE) & 0.7883 (80.72% @ $0.21/1K) in RouterArena CI (PR #169 open) • Zero E2E QA drop from 100K → 1M tokens (72.86%) on ICLR 2026 BEAM, beating cloud past. dev on Contradiction Resolution (4/4 splits) & 10M Multi-Session Reasoning
1
1
156
I ran all 400/400 ICLR 2026 BEAM (100K) memory probes inside an air-gapped 96GB Blackwell GPU—zero cloud APIs, dual cross-family judges, 5 SHA-256 anti-gaming gates:
• 92.4% 8K Recall (+5.3pp R@10 w/ 8K Arctic)
• 98.1% Contradiction Resolution (120B QLoRA)
krusch.dev/articles/iclr-202…
1
1
142